North American time on September 30, 2022 (Beijing time on October 1, 2022), Tesla held the 2022 AI Day (AI DAY) event at the headquarters in Alto, Palo, California, USA. Tesla CEO Musk led the Tesla team to showcase the prototype of the humanoid robot Optimus, and introduce the latest progress of the fully autonomous driving system FSD, supercomputing platform Dojo, etc.
This issue of "long-termism", choose Tesla's 2022 AI daily record, press conference minutes, on-site Q&A minutes, rubble village man, smart things, car things, electric planet News, Wall Street News released, Liuhe Commercial Research and Selection and Compilation, Share with everyone, enjoy!
Text:
Full text 23,848 words
expected to read for 48 minutes
Tesla 2022 AI Day Record
Word Count: 7,202 Words
Source: Zhidongxi, Car Games, Electric Planet News, Wall Street News
North American time On September 30, 2022, Tesla held an AI Day event. Unlike the previous Musk's lecture mode, on this AI Day, Musk briefly introduced the humanoid robot Optimus that highlighted it, followed by the heads of each business to give a specific explanation. The
activity began. The humanoid robot Optimus Optimus prototype, which had no decorations around, slowly walked onto the stage and completed walking, turning, waving and other actions. Tesla also plays a video of Optimus Optimus, showing that Optimus Optimus can complete transporting items, watering flowers and other activities. According to Musk's vision, Optimus Optimus can be used in the future for homes, cooking, mowing lawns, taking care of the elderly, and even becoming a human partner or companion.
Optimus Prime Optimus appears
Tesla 2021 AI Day, Optimus Prime Optimus is just a concept. In the past year, Tesla has successfully launched a prototype that can walk and complete multiple actions. In order to enable Optimus Optimus to be released smoothly, Musk postponed Tesla's 2022 AI Day for more than 40 days from the original August 19, 2022.
Musk attaches great importance to Optimus Optimus in the future, saying that in the future, Optimus Optimus will be more important than cars and autonomous driving. In the future, Tesla will produce millions of robots, and the price will be cheaper than Tesla car , which may be less than 20,000 US dollars, and consumers will be able to buy it in the next 3 to 5 years. In terms of autonomous driving, Tesla has been improving its technology and products. Currently, the FSD Beta version of the fully autonomous driving system has been used by 160,000 users, and it is expected to have global promotion capabilities by the end of 2022.
Tesla's self-developed supercomputing platform Dojo. Currently, this product is helping Tesla achieve more achievements in autonomous driving, robots, etc.
At the entire press conference, several Tesla engineers introduced Optimus Optimus, FSD, Dojo, etc. in detail software and hardware. When introducing each product, they talked about hoping that more AI, hardware and other talents would join Tesla. Musk also said that the purpose of holding this event is to attract more AI talents to join Tesla and make better products.
Optimus Prime Optimus prototype was unveiled. The expected price is no more than US$20,000
The press conference began. Musk showed the Tesla humanoid robot Optimus Prime Optimus prototype. From the appearance, the robot that debuted first was indeed very prototype, with wires exposed all over the body without any decoration. Optimus Optimus can independently complete walking, wave to users, and even dance.
Optimus Optimus Dancing
Musk said that Optimus Optimus can do more things, but due to the stage, it can only show these. Tesla plays Optimus Optimus video live. In addition to walking around, Optimus Optimus can also complete transportation of items and watering flowers.
Optimus Optimus transports things
In the factory, Optimus Optimus can take a long strip of objects out of the workbench and then neatly discharge them into a box containing the same objects. Judging from the renderings from Optimus's own perspective, it can distinguish different objects in the real world by colors, such as seeing the handheld long strips as purple, the workbench as yellow, etc.
Optimus Optimus transports goods
Tesla then displays a real version of Optimus Optimus that is closer to real people. Its appearance is similar to the 2021 AI Sun Display model, with a human appearance and higher degree of freedom. The real-time Optimus version of Optimus can provide more services, such as free movement of fingers, operating more tools, holding some tools with your right hand, and even doing some repetitive work in the factory. Unlike the initial version, the real version of Optimus Optimus did not walk around the site, but was carried by staff throughout the process.
Musk said that the Tesla robot team has worked almost 7 days a week and worked more than ten hours a day in the past year, and finally successfully displayed the Optimus Prime Optimus prototype. It took only more than 6 months from the launch of the first-generation R&D platform in February 2022 to the launch of the prototype. The Tesla team has done a lot of work, and Optimus Optimus is still in its early stages and can do it better in the future.
real machine version Optimus Prime Optimus
Musk said that Optimus Prime Optimus robot project represents the mission expansion of Tesla. Optimus Optimus is designed based on the human body and will have the ability to communicate and hope that its behavior will be as close as possible to the human behavior. The future full-body version of Optimus Optimus has a weight of 73kg. It uses 100W of electricity when sitting in a meditation, 500W of electricity when walking quickly, and has more than 200 degrees of freedom in the whole body and 27 degrees of freedom in the hands.
Musk predicts that Optimus Optimus can be produced at low cost, with a future output of millions of units, with a cost of less than US$20,000. Compared with similar robot products on the market, the cost will be significantly reduced.
Musk said that robots can reduce labor costs and make social development more motivated. There will be no poverty in the future. Humans can freely choose the type of work and no longer have to do physical work. They can participate more in mental work. I hope that robots can bring more help to humans more safely. Self-driving cars have a huge impact on the world, increasing transportation productivity by at least half an order of magnitude. In the future, robots may have several orders of magnitude improvements.
Tesla's first generation robot has been repeatedly verified from concept, design, analysis, establishment, and optimization. For this new species, having lower cost and higher work efficiency is the key to verifying whether the product can truly be launched into the market.
From the more detailed functions of Optimus Optimus watering flowers and moving things, Tesla insists on constantly evolving towards anthropomorphic robots, from hand movements, gait adjustments, control systems, etc., relying on Tesla's strong technical accumulation in the automotive field, infrastructure and supply chain capabilities, plus Musk's ambitions and strong action power in the field of humanoid robots, bringing unlimited possibilities to the robot industry. After Musk briefly introduced it, the Tesla robot team introduced Optimus Optimus. Starting from Tesla's 2021 AI Day, Optimus has evolved three times so far, and finally achieved results.
Optimus Prime Optimus design is based on the vehicle design process accumulated by Tesla before. Tesla technicians say that cars are wheeled robots, and Optimus Optimus just stands up the car to some extent.
From the Optimus Optimus concept diagram, the entire robot contains 28 electric drives (orange) and battery and control module (blue).
Humans can eat a small amount of food to maintain energy. In order to reduce the energy consumption of robots, Tesla minimizes the energy consumption of humanoid robots when they are idle. Just press the switch to adjust it to be in a low battery or normal working state.
Optimus Prime Optimus power system is integrated into the upper body of the robot, and contains a battery pack with a capacity of 2.3kWh, with a working voltage of 52V, and can run for a whole day after charging. What’s unique about this battery pack is that it integrates all batteries, sensors, etc. using automobiles and energy products. This design is a reference to Tesla's automotive design plan, hoping to reduce more wiring harnesses and focus distribution and calculations on the center of the torso.
Optimus Optimus uses a single Tesla self-developed SoC chip, which supports LTE 4G connection, which is different from the dual-chip autonomous driving solution that Tesla uses in cars. Unlike cars, Optimus Optimus needs to process visual data to respond quickly, which is based on a variety of sensory inputs and communications. Therefore, it is equipped with modules such as radio connection and audio support, which have the characteristics of protecting the robot body and human safety.
Optimus Optimus core components display
In terms of action, Optimus Optimus Optimus absorbs Tesla's automotive power system design experience, and the team analyzes what actions Optimus Optimus needs to take, including walking, going up and downstairs, etc. Then, first analyzes the dynamic data of Optimus Optimus walking, and then analyzes the time, energy consumption, and trajectory required for these movements, and designs the robot joints and actuators based on these data. In terms of security of
, Tesla has made targeted designs to protect robots, R&D personnel optimize their infrastructure. For example, when Optimus Optimus falls, it will not damage the transmission and the arm. After all, the robot repair cost is very high.
Optimus can protect itself when falling
Tesla R&D personnel use the same underlying technology as Tesla cars, making pressure in all components of Optimus Optimus, making it easier to control its walking and not appear stiff.
Optimus Optimus Optimus Walking Pose Simulation
Taking the knee as an example, Optimus Optimus was designed according to the real human knee structure. The R&D personnel simulated the human knee and the force they were exposed during movement, learning how to use less force on the robot knee, allowing it to achieve better force control, and allowing the relevant structure to be tightly wrapped around the knee.
Optimus Prime Simulates human bone structure
Cars and robots have many similarities in power design, so Tesla's experience in power design can be applied to robots. The car drive unit is to accelerate the car. The robot has 28 joint drives, which is not much different from the car drives. However, the tasks that humanoid robots have to do are more complex and require walking or climbing. Therefore, the R&D personnel use models to generate the robot's connection torque speed trajectory, and then enter its optimization model to run.
Cars and Robot Driver Comparison
Robots need to rotate. According to the torque speed trajectory and joint driver efficiency diagram, the energy consumption generated along the track, the cumulative energy of the execution tasks and running time, can define samples of specific actuators and send them to the cloud. This will reduce the time it takes to produce millions of joint drivers.
However, each joint of the robot is specific, and 28 joints require customized specifications. Therefore, R&D personnel need to reduce customized actuator designs, analyze and upload them to the cloud for common research.
Optimus Optimus has 6 types of actuators, including 3 different specifications of servo , 3 different specifications of traction actuators similar to weight scales, etc. Tested within the limits, its joint drive lifts a half-ton piano, which is also a humanoid robot must have functions.
Optimus Prime Optimus joint driver lifts a half-ton piano
Human hands can move at a speed of 300 degrees per second, and have thousands of tactile sensors . Another focus of Optimus Optimus is the hand. Tesla hopes that Optimus Optimus' hands can be as flexible as humans, able to grasp objects, operate, have sensors for perception, etc.
Tesla R&D team also draws inspiration from biology when designing Optimus Optimus's hands. Through 6 actuators, Optimus Optimus can achieve 11 degrees of freedom movement, hold objects weighing 20 pounds (9.1 kg), operate some equipment, or grab small objects, etc.
The adaptation mechanism of the hand is very complex. Humans can recognize the position of the hand in space, which is also the key to its ability to grab objects. Therefore, Tesla is currently conducting corresponding tests.
Optimus Optimus uses the core sensor as the camera, which is similar to the camera used in Tesla's FSD system. Tesla currently collects a lot of data to train Optimus Optimus. In terms of action, Tesla Automobile also uses neural network "occupies the network" to identify the actionable area. After sensing and analyzing the external environment, the software will plan the travel trajectory, and then plan each footing point according to the trajectory, and finally perform the moving action through the actuator.
Optimus Optimus uses the "occupying network" to identify mobile areas
Tesla R&D personnel demonstrate the growth trajectory of Optimus Optimus Optimus's action capability. In April 2022, the first step was taken; in July 2022, the application of a robot unlocking the pelvis to maintain balance; in August 2022, the arms began to work; in September 2022, Optimus Prime Optimus toes also came in handy.
As the humanoid robot slowly utilizes more joints and trains it, the movement speed is significantly improved. Optimus Optimus is currently walking very slowly, not reaching the 5 mph (about 8 km/h) level claimed when it was first released in 2021.
Optimus Prime Optimus action capability growth trajectory
From the perspective of human walking process, it is important for robots to have a physical self-awareness, such as being able to realize the length of their limbs, how to maintain balance, and how to coordinate the movement of their limbs. It's easy for humans to do this, but it's hard for robots.
Optimus Another focus is to keep it upright and not fall to the ground easily. Through the motion planning and control stack, R&D personnel generate robot kinematics models, and then generate the required paths to the underlying platform, allowing the entire system to refer to its trajectory for training. In the Optimus Prime Optimus simulation system, the lines are planned according to its desired path, and the interconnected trajectories are constantly added. Each footing point is planned according to the trajectory, and then the actuator is allowed to execute it, so as to ensure the balance of the robot during walking.
Optimus Prime Walking Upright
In sports training, the motion planning method is an ideal situation, but in fact it is not realistic to put it in the real world.A few key actions are missing in the process, and placing them into the real world can have an impact on model dynamics, especially systems like bipedal dynamics.
R&D personnel use corresponding sensors and observe in the real world to solve robot control problems, use robot pelvic position, center of gravity position, etc. to track the path of the robot in the studio, build a better robot model and correct robot behavior based on actual conditions.
Tesla hopes to make Optimus more flexible in the future, go further from the prototype function, so that it can improve in all aspects, and have better navigation, action and other capabilities.
To solve the problem of humanoid robots manipulating the real object world while observing, R&D personnel divide this problem into two steps, first generate a natural motion reference system, and then optimize and promote it.
Assuming that someone demonstrates a certain action, the R&D personnel set keyframes to its action through video and map it to the robot. This requires only a demonstration video, and R&D personnel can promote it to the robot's realistic manipulation application. This can solve the problems of where the robot should put his hands when grabbing an object, how to carry and place it.
Optimus Optimus performs crawling actions by simulating real people
FSD is progressing smoothly. It is expected to have global promotion capabilities by the end of 2022
In terms of autonomous driving, the Tesla team first introduced the FSD situation of the fully autonomous driving system. FSD Beta, with 2,000 customers participating in the test in 2021, expanding to 160,000 customers in 2022, achieving 80 times growth. Since 2022, Tesla has trained 75,000 neural network models and launched 35 FSD version updates on this basis. In terms of the technical architecture of autonomous driving, Tesla's approach is to first use an automated data annotation system to automatically label the collected data, then process the data to train the neural network, and then deploy the AI model to the FSD computer, calculate the perception results of the external environment through calculations, and then plan the vehicle's driving route. In terms of technology display of
, Tesla first demonstrated its unprotected left-turning ability. For example, when turning left, there are pedestrians and other vehicles. After taking into account the driving trajectories of different traffic participants, Tesla calculated the most suitable driving trajectory.
Tesla FSD no protection left turn capability
Tesla uses the name interactive search technology. First, start with visual perception, perceive traffic participants, then infer their driving trajectory, then generate several strategies, and finally select the best driving trajectory.
Tesla interactive search technology architecture
It should be noted here that if there are more and more external targets, the amount of calculations will become larger and larger. Tesla perceives the outside world by looking around the camera, generates a 3D environment, and finds the driving areas through the "occupying the network" to identify which obstacles are. When working in
FSD, the camera image is calibrated first, the images are integrated together, the images are formed, the data is extracted, and the data is input into the neural network. Through the corresponding algorithm, spatial features are constructed. After
only generates 3D space, there is no precise location of various objects, and it is still impossible to plan the path. Therefore, Tesla analyzes key features and calculates the location data.
Tesla FSD builds a 3D space based on visual data
Tesla team accumulates a lot of video clips during daily driving, and each video has many frame images.Tesla needs 1.4 billion images to train a neural network, and needs to use 100,000 GPU working hours (1 GPU 1 hour) to work. The computing power is huge and supercomputers and AI accelerators are required. This is also an important reason why Tesla's self-developed supercomputing platform Dojo. Based on Dojo, Tesla can increase the training speed of neural networks by 30%. In terms of predicting the behavior of other traffic participants, Tesla's approach is to first enter the RegNet network ( convolutional neural network developed by Meta's FAIR laboratory), and then the processed data enters the Transformer model ( deep learning model developed by Google ). There may be 1 billion parameters to optimize them together. To achieve maximum computing power and minimize latency.
Tesla cars will generate a large amount of data during operation, and these data also need to be marked. In terms of data annotation, Tesla first tried manual annotation, which was time-consuming and labor-intensive; then considered the supplier cooperation method. Judging from the final results, both timeliness and quality are not very good. Tesla needs very efficient and scalable annotations.
Tesla currently adopts the standard method of human-machine cooperation, including both human and machine annotations. Overall, machine labeling is more efficient, and the machine has a 30-minute workload, which may take a long time for humans, so Tesla is building an automatic labeling system. Through efficient annotation, space-time fragments in the real world are converted into usable data, making FSD more intelligent and efficient. Automatically labeling of data also needs to be sorted out. Tesla did not invest much energy in this area before, and now many engineers are doing this.
Tesla trains the model based on automatic labeling data
simulation system is a very important part of the autonomous driving system, which can improve the vehicle's ability to respond to long-tail scenarios. Tesla has developed a scene generator that can generate a scene in as fast as 5 minutes. Compared with the traditional method, the speed is 1,000 times higher. It can also scan real objects to project onto the screen, simulate signal lights, stop signs, etc., and be as close to the real world as possible. This is of great significance to model training.
Tesla R&D Scene Generator
Currently, Tesla data sets are partly from the information sent back by the fleet and partly from the simulation data, which can make judgments on the scene more convenient. Through the data engine, neural networks can be made more realistic, bring more certainty to FSD, and solve uncertainty in the real world. If a car turns at an intersection, it is necessary to determine whether the parking vehicle is parked or is driving slowly. Just create more networks for evaluation to solve this scenario problem.
Regarding the promotion of Tesla's FSD beta version, Tesla said that by the end of 2022, it will have the ability to promote FSD globally. In addition to North America, Tesla also needs to communicate with local regulatory authorities. In some countries and regions, the regulatory policies for autonomous driving are still lagging behind.
supercomputing platform Dojo is constantly iterating to ensure Tesla's computing power needs
Tesla team has mentioned the supercomputing platform Dojo many times when introducing Optimus Prime Optimus robot and fully autonomous driving FSD.
Tesla 2021 AI Day, the first Tesla's first AI training chip Dojo D1 is displayed for the first time, and the supercomputer system ExaPOD built on this chip can be used to perform AI training tasks and provide support for the huge image processing needs of on-road vehicles.
Tesla currently has a large supercomputing platform based on Nvidia GPU, and a data center that stores 30PB video materials. It is currently developing a supercomputing platform based on Dojo chips.
Tesla uses a set of pictures to show the key nodes of Dojo on the supercomputing platform in the past two years, from the delivery of a customized refrigerant distribution unit CDU, to the installation of the first integrated Dojo cabinet, and then to the 2.2MW unit load test.
Supercomputing platform Dojo Key node
Tesla tries to optimize the scalability of Dojo design and overcome challenges in a fast trial and error manner. Dojo has features such as a single scalable computing plane, global addressing fast memory, unified high bandwidth + low latency, etc.
Tesla technical engineer, especially when talking about the Dojo voltage regulation module. In the past two years, the voltage regulation module has been updated in 14 versions, with high performance, high density, complex integration and other characteristics.
supercomputing platform Dojo voltage regulation module
thermal expansion coefficient CTE is very important. Tesla cooperates with suppliers to provide power solutions to reduce the thermal expansion coefficient CTE of the supercomputing platform Dojo over 50%, making Dojo performance reach 3 times the initial version.
Tesla team demonstrates the use of Dojo to help AI drawing software Stable Diffusion to draw Tesla electric pickup Cybertruck to exercise images on Mars.
Cybertruck on Mars picture
According to reports, only 4 Dojo cabinets can replace 72 GPU racks of 4,000 GPUs. Dojo can reduce the working hours usually take several months to 1 week.
Tesla's self-developed D1 chip also plays an important role in the supercomputing platform Dojo. The D1 chip adopts TSMC's 7nm process process , and 50 billion transistors are distributed over an area of 645mm². The peak computing power reaches 362 TFLOPS (BF16/CFP8 accuracy, BF16 and CFP8 are floating point calculation format standards), and the thermal design power consumption TDP does not exceed 400W.
is based on D1 chip. Tesla launched a system-level solution on wafers. By applying TSMC InFO_SoW packaging technology, 25 D1 chips are used to form a training module Training Tile. The peak computing power of one training module reaches 9 PFLOPS (with BF16/CFP8 accuracy), and includes computing, I/O (input/output), power, liquid cooling and other modules. It adopts centralized power supply and heat dissipation design, and the heat dissipation power reaches 15kW.
supercomputing platform Dojo's training module Training Tile
Dojo system tray System Tray has high-speed connection, dense integration and other characteristics. The peak computing power can reach 54 TFLOPS (with BF16/CFP8 accuracy), and the power consumption is 100+kW.
Dojo interface processor is a PCIe card with high bandwidth memory, using Tesla's own TTP interface. Tesla transmission protocol TTP can bridge to standard Ethernet, and TTPOE can convert standard Ethernet to Z-plane topology , with high Z-plane topology connectivity.
Tesla 2021 AI Day, Dojo development has ushered in a series of milestones, including installing the first Dojo cabinet and conducting 2.2mW load testing. Tesla is currently promoting Dojo R&D at the speed of creating a training module every day.
supercomputing platform Dojo cabinet
Tesla will form a Dojo supercomputer with 60 training modules, with 1,500 D1 chips and more than 530,000 training nodes. In theory, there is no upper limit for Dojo performance expansion and can be expanded infinitely.
In actual applications, Tesla will form 120 training modules to form a supercomputer system ExaPOD, with 3,000 D1 chips and more than 1 million training nodes, with peak computing power exceeding Level E, reaching 1.1 EFLOPS (with BF16/CFP8 accuracy), and also provide 13TB memory and 1.3TB cache capability.
Tesla team said that computing power is the foundation of fully autonomous driving, and high-efficiency chips can better serve fully autonomous driving. What Tesla needs to do is to make Dojo the strongest supercomputer system in AI training.
Tesla does not limit Dojo design. It can provide a very large and efficient system, breaking the physical boundaries of traditional integration from the hardware, and making the hardware more efficient on the compiler. As long as physical conditions allow, Tesla can continuously break through the limit.
Tesla announced that it is expected to deploy the first ExaPOD in Q1 2023. After successful construction, ExaPOD will become one of the most powerful supercomputers in the world.
Tesla said that with the exclusive compiler, DOJO training delay can be achieved at least 1/50 of the same-scale GPU. Tesla's goal is that by mass production in Q1 2023, DOJO can achieve 4.4 times the single-chip training speed of Nvidia A100, and even lower energy consumption and cost.
Tesla 2022 AI Day Press Conference Minutes
Word count: 5,656 words
Source: Ruble Village Man
Tesla 2022 AI Day, after Musk briefly introduced the humanoid robot Optimus Optimus, the heads of Tesla's business gave a specific explanation.
humanoid robot Optimus
Musk: Welcome to Tesla AI Day 2022. We have some very exciting content to show you and I think it will impress you.
For Optimus Optimus Robot, I do want to set expectations. In 2021, it was just a guy wearing a robot suit and now we have made great progress. Compared to 2021, it will be impressive. We will talk about our progress in AI, autonomous driving, and Dojo.
Should we let the robot appear?
Kate: Before that, we had a small tip, which is actually the first time we tried this robot without any fallback support. Cranes, mechanical devices, cables, etc., nothing.
Musk: We will show you videos of robots doing other things.
Milan: We want to show some of the progress made around robots in the past few months, which can move around and dance.
This is just a small start. You can see that the autonomous driving neural network is running as it is. We are just retraining the robot directly on the new platform.
Musk: When you see the rendered view, that is the world that the robot sees. It can identify objects very clearly, such as the ones it should pick up.
Milan: We use the same process as the autonomous driving system to collect data and train neural networks. For robots, we also deploy this way.
This is an example to further demonstrate the robot's upper body function. We do want to improve this part of the functionality in the next few months.
Kate: That is not the only content we want to show today.
Musk: The robot you just saw is called Wasp Bumble C. It is a development version of the robot, which uses a semi-off-the-shelf drive.
In fact, we have gone a step further and our team has done an amazing job. In fact, we have an Optimus Optimus robot that uses drives, battery packs, control systems and everything else that is completely designed and produced by Tesla. It can't walk, but it will be within a few weeks.
We want to show this robot, but in fact, it is quite close to the production standard. We want to show all the things it can do, let's ask for a robot.
You can see Optimus Optimus, the degree of freedom it has is what we hope the first mass production machine can have. That is, the ability to move all fingers independently, the thumb has two degrees of freedom. It has an opposing thumb, a left hand and a right hand, and can operate tools and do useful things.
Our goal is to produce useful humanoid robots as soon as possible.When designing it, we adopt the same principle as designing cars, which is designed for production. This is the only way to produce robots with high output, low cost and high reliability.
Optimus Optimus design goal is to produce millions of robots with extremely strong capabilities and extremely high production. Optimus Optimus is expected to cost much less than a car, probably less than $20,000, which is my guess.
Its potential is unfathomable. You can say, what is economy? The economy is the number of entities that carry out production multiplied by productivity and the population multiplied by per capita output. Once there is no limit on the population, what the economy actually means will be less clear and the economy will be close to infinity.
This means a rich future, a poor future. At that time you could have any product and service you wanted, which was indeed a fundamental transformation of human civilization as we know it.
It is very important that corporate entities that turn this ideal into reality need to accept the reasonable impact brought by the public. I think Tesla's structure is very ideal.
Kate: You have seen several robots today, let's quickly review the timeline.
appears for your robots who perform talents. We will complete the production within 6 months, and carry out software integration and hardware upgrades within 1 month thereafter.
At the same time, we are also designing the next generation of robots, this robot here. This guy, rooted in the vehicle design process, is leveraging all of this experience already.
Repeat, we adopt the basics of vehicle design, from concept to design and analysis, and then build and verify. In this process, we will optimize cost and efficiency, and ultimately these are key indicators for the product to scale.
Inside the torso, we install the battery pack with a capacity of 2.3kWh, which is a perfect configuration for all day use. Next is the brain, which is not on the head, but is very close to it, in the torso, we install the central computer.
Tesla has equipped every car with an FSD system. We want to use the hardware and software of the autonomous driving system to develop a humanoid robot platform. But because its needs are different from appearance factors, we must first make changes.
It requires doing everything the human brain does, including processing visual data, making instant decisions based on multi-sensor input, and communication. To support communication, it is equipped with wireless connection and audio support. It also has hardware-level security features, which is important for protecting robots and people around them.
Nilegen: Can we use our capabilities and methods on the automotive side to influence robots?
Since we have collision software, we can use the same software to make it fall down. The purpose of this is to ensure that even if the robot falls, it will only be damaged on the surface. Of course, it is best not to fall. We hope it dusts and continues to complete the task.
drive can lift a concert grand piano weighing half a ton and 9 feet long. The design of the robot hand is inspired by biology and has 5 fingers, which are driven by metal tendons, flexible and strong, and have the ability to grasp a large range of force. It is also optimized for precise grasping of very thin and small objects.
Milan:We show all this cool content in the video, all of which are done in just a few months. Thanks to our magical progress on autonomous driving systems over the past few years, most of the components can be easily ported to a robotic environment.
You can think of it, this is just turning the robot on the wheel into a robot with long legs, some of which are very similar, others require us to put in more work.
For example, our computer vision neural network is directly transplanted from the autonomous driving system to the robot environment. We are also working hard to find ways to use neural radiation fields to improve these "occupying networks" and obtain a good volume rendering effect for the surrounding environment of the robot. For example, a robot interprets what it needs to interact with.
Another interesting question is, in indoor environments, there is no GPS signal in most cases, how to make it navigate to the destination?
We have been training more neural networks to identify high-frequency features and key points in the robot camera to obtain images, and track across frames and time in the robot navigation environment. We use these points to better estimate robot poses and trajectories in the work environment.
This is a video that demonstrates the code to run motion control in an autonomous driving system simulator and demonstrates the evolution of the robot's walking ability. As can be seen, it walked quite slowly when it started in April 2022. Over the past few months, it has begun to accelerate as more joints are unlocked with more advanced technologies such as arm balance.
Hopefully so far, you have a good understanding of our work in the past few months. We started implementing a usable robot, but it was far from practical. There is still a long and exciting road ahead of us.
I think in the next few weeks, we need to complete the first task, to get Optimus Optimus to at least reach or even surpass the Wasp level, which is the robot prototype you just saw.
We will also start to focus on real usage scenarios in one of our factories, committed to truly solving problems, and thoroughly consolidate all the elements of deploying this product to the real world, including the indoor navigation, elegant comprehensive management, and even providing services, and all the components required for large-scale production.
I don't know what you think, but after seeing what we showcased, I'm sure we can achieve this goal in the next few months or years, turn this product into reality, and change the entire economy.
I thank the entire Optimus Optimus team for working hard over the past few months and I think they work very well. All of this was completed in just 6 to 8 months, thank you very much.
Automatic driving/FSD Beta version
Ashok:
In the past year, we have trained 75,000 neural network models and completed one model training every 8 minutes. We evaluated on a large computer cluster and released 281 of them, which indeed improved the performance of the car.
This innovation speed appears in all aspects of the entire technology stack, including planning software, infrastructure, tools, etc. Everything is developing to a higher level.
We take this intersection scenario as an example to explore how autonomous driving systems plan and make decisions.
We walk from the side path to the intersection and have to make way for all vehicles crossing the road. Just as we were about to enter the intersection, pedestrians on the other side of the intersection decided not to walk the zebra crossing and cross the road. Now we must give way to this pedestrian and to the vehicles coming to the right. We also need to understand the relationship between pedestrians and vehicles on the other side of the intersection. We need to quickly judge the dependence relationship between a large number of objects.
Humans are very good at this. We see a scenario that understands all possible interactions, evaluates the most likely interactions, and usually ultimately chooses a reasonable judgment. But the frame can also be expanded to the objects behind the occlusion.
We use video sources from 8 cameras to generate three-dimensional occupation information of the surrounding world. The blue part here, corresponding to what we call the "visible area", is basically blocked by the first obstruction you see in the scene. We use the model to generate "ghost objects" in the "visible area". If you correctly model the generation area and state transitions of "ghost objects", if you adjust the control reaction as a function of the possibility of existence, you can extract some very good human-like behaviors.
Phil: "occupy the network" receives all 8 of our camera video streams as input, and directly generates a unified occupancy rate in the vector space. For each three-dimensional position around our car, the probability that that position will be occupied is predicted.
Tim: Let's talk about the infrastructure for training. We have watched four or five videos, and the number of video clips I think about and care about is much larger than that.
We just watched the "occupying network" introduced by Phil. This video alone requires 1.4 billion images to train the network you just saw. If you have 100,000 GPUs, you need 1 hour; but if you only have 1 GPU, you need 100,000 hours. This training task takes a length of time, not something you can afford.
We hope to publish it faster, which means that we need parallel processing, we need greater computing power, which means we need supercomputers.
This is why we built 3 supercomputers within the company, including 14,000 GPUs. We use 10,000 of these GPUs for training, and about 4,000 GPUs are used for automatic annotation.
I can keep talking. I just briefly introduced our two internal projects, which are actually just part of the huge project to optimize our internal computing power.
Through the accumulation and integration of all these optimizations, we now train "occupy the network" twice as fast as it is because its efficiency is doubled. If we add more computing power and use parallel computing, we can complete the training in a few hours, not a few days.
John: I am the head of the visual team of the autonomous driving system. Today I want to introduce two topics to you. First, how do we predict the lane; second, how do we predict the future behavior of other objects on the road.
What we obtain through this lane detection network is a series of lane connectivity, which is directly calculated and output by the network. There are no additional steps here, and there is no need to apply intensive predictions to decentralized predictions, which is the direct output of an unfiltered network.
Above I discuss some content about lane detection. I will briefly discuss how to model and predict the future paths of other objects. I want to quickly show two examples. In the video on the right of
, there is a car running a red light and turning in front of us. Our approach to this situation is to make a series of short-term cycles of future trajectory predictions for all objects. We can use these results to predict possible dangers and use braking, steering and other behaviors to avoid collisions.
Overall, the visual technology stack of the autonomous driving system predicts not only the geometric motion parameters of the surrounding world, but also the rich semantics, thereby achieving safe driving similar to humans.
Jaegan: I will talk about automatic annotation. We have several automatic annotation frameworks that support various types of networks. Today I would like to focus on this excellent lane network.
This network is easy to expand as long as we have enough computing power and itinerary data. In this scenario, about 50 trips are automatically marked, some of which are shown here, 50 trips from different vehicles.This is the process of our capture and transforming space-time fragments of the world into network supervision.
David: Take the simulation scene played behind me as an example. It takes an artist 2 weeks to complete the design. This is too slow for us.
I will talk about using Jaegan's automatic benchmark annotation and some brand new tools. We can generate this scene in just 5 minutes and processically and many similar scenes. This speed is amazing, 1,000 times faster than before.
This method is for scale and scale, be prepared. As you can see on the map behind, we can easily generate most urban streets in San Francisco without spending months or even years, just one person working for 2 weeks.
We reviewed that because we generate all fragment data sets through benchmark data, including all the complex situations in the real world, we can combine process-based vision and various changes in traffic conditions to create infinite target data for online learning.
Kate: This data engine framework is suitable for all signals, whether it is three-dimensional multi-camera video, whether the data is manual standard, automatic annotation, or simulated data, whether it is offline model or online model.
Tesla can optimize on a large scale, thanks to the advantages of the fleet, the infrastructure built by our terminal team, and the labeling resources provided for our network. To train all this data, we need a lot of computing power.
supercomputing platform Dojo
Pete: I am often asked why a car company wants to build a supercomputer for training?
raises this question, but it is still a fundamental misunderstanding of the essence of Tesla. Essentially, Tesla is a hard-core technology company.
Yaji:
Since then, the entire team has been working hard and committed to large-scale deployment. Now we have made amazing progress, and we have achieved many milestones throughout the process, and we have encountered many unexpected challenges. It is our philosophy of "quick trial and error" that allows us to break through our own limits.
starts with our custom D1 chip uniform nodes, connect them to our fully integrated training module, and then finally seamlessly connect them across the boundaries of the cabinet to form the supercomputing platform Dojo.
In short, one ExaPOD can accommodate two Dojos, and the overall computing power reaches 1.1 EFLOPS. In computing history, this degree of technology and integration have only appeared a few times.
Rajeef: This operation takes only 5 microseconds on 25 Dojo chips, while the same operation takes 150 microseconds on 24 GPUs, which is an order of magnitude improvement compared to the GPU. How are the two networks performing? We are going to see the results, which are calculated on multi-chip GPU and Dojo systems, but are both normalized to single chip values.
On our automatic labeling network, the previous generation of VRMS software running on the current hardware can surpass the performance of Nvidia A100; on our production hardware, we can run our newer VRMS software to achieve twice the throughput of A100. Our model shows that with some key compiler optimizations, we can achieve more than 3 times the performance of the A100.
We have seen a bigger leap in "occupying the network". Using our production hardware, we have almost achieved a 3x performance improvement, and there is still more room for improvement.
used to take more than a month to train the network, but now it only takes less than a week.We start with hardware design, breaking through the boundaries of traditional integration and serving our vision of a single giant accelerator. We have seen how to build a compiler on this hardware.
Dojo performance is proven through these complex real-world networks. We also know what our first large-scale deployment should be, our high computing strength automatic annotation network.
Today these networks use 4,000 GPUs on 72 GPU racks. With our intensive computing power and high performance, we hope to provide the same computing throughput with just 4 Dojo cabinets. These 4 Dojo cabinets will be part of the first ExaPOD, which we plan to deploy in Q1 2023, more than twice the existing automatic annotation capability of Tesla.
The first ExaPOD is part of the 7 ExaPODs we plan to build in Palo Alto, California, USA, just across the wall. We have 1 of the ExaPOD display cabinet for everyone to watch.
Conclusion
Musk: We really want to show the depth and breadth of Tesla technology, as well as AI, computing hardware, robots, drives, etc.
We strive to change people's perception of Tesla. Many people think that we are just a car company and we only make cool cars. Most people don’t know that Tesla can be said to be a global leader in AI, hardware and software.
We are building arguably the first, possibly the most radical computer architecture since the Cray-1 supercomputer (the fastest supercomputer in the world from 1976 to 1982).
If you are committed to developing the world's most advanced technology and truly influence the world in a positive way, joining Tesla is the right thing to do.
Tesla 2022 AI Day Live Q&A Minutes
Word Count: 10,990 words
Source: Ruli Village
Tesla 2022 AI Day, Musk led the Tesla team to introduce Optimus Prime Optimus prototype, fully autonomous driving system FSD, and supercomputing platform Dojo, and answered 23 sets of questions raised by the audience.
Musk: I hope we will tell enough details and accept the question now.
Q1: Optimus Optimus left a deep impression on me. I wonder why my hands use rope driving method? Why do you choose rope driving methods for your hands? Because the mechanical tendon is not very durable. Also why spring-loaded models?
Mike: This is a good question. When evaluating any kind of drive scheme, whether it is a rope drive system or a connecting rod system, there must be some choice.
The main reason we chose the rope drive system is that first we investigated some synthetic tendons and found that the strength of metal marine cables is much higher. One of the advantages of these cables is that they can reduce energy consumption very well.
We want to produce a large number of hands. When mass production is carried out, many parts and many small connecting rod devices will eventually become a problem. An important reason why rope drive is better than connecting rods is that it can eliminate gaps. The essence of gap elimination is to keep your fingers from lagging when they move. The main benefit of the
spring loading model is that it allows us to actively open our hands. We don't need to use two drivers to drive the fingers to close and open, we have the ability to let the mechanical tendons drive them to close and then the spring is passively elongated.
This point can also be seen in our hands. We have the ability to actively bend our fingers and also have the ability to stretch our fingers.
Musk: Our purpose of designing Optimus Optimus is to realize the robot that plays its greatest role as soon as possible.
There are many ways to solve various problems of humanoid robots. We may not have found the correct answer in all technical solutions. I should say that we are open to the technological solutions you see evolve over time, and they are not static.
But we have to make a choice, we want to choose something that will allow us to achieve production as soon as possible, as I said, it can work as soon as possible. We strive to follow our goals to produce useful robots at the fastest speed and can be mass-produced.
We will conduct internal testing of this robot in the Tesla factory to see how effective it is. We need to complete a closed loop in reality to ensure that the robot is useful.
We are confident that we can use the hands we are currently designed to achieve this goal, but it is certain that the hand design will have the second and third editions. Over time, we may make considerable changes to the robot structure.
Q2: Optimus Prime Optimus robot is indeed impressive, and bipedal robots are indeed difficult. But I noticed that you may lack recognition of the spiritual value of human beings in your plan. I wonder if Optimus Optimus has personality and can be amused by us when folding clothes for us?
Musk: We hope there is a really interesting version of Optimus. Optimus Optimus can be utilitarian, can complete tasks, or play with you like friends and partners.
I believe people will come up with various creative uses of this robot. Once the core intelligence and drive problems are solved, you can put on various vests on the robot and replace the skin of the robot in different ways.
I believe people will come up with various interesting versions of Optimus Optimus.
Q3: I want to know if there is any behavior equivalent to manual intervention for Optimus Optimus. It seems that it is important to mark moments when humans and machines have different judgments. For humanoid robots, this may also be an ideal source of information?
Ashok:I think we will have multiple ways to remotely manipulate the robot and intervene when it goes wrong, especially when we are training the robot.
We hope to design it in some way that if it is going to hit something we can press a button and it will stop without crushing your hand or something, which are intervening data.
We can also learn a lot from our simulation system, which can check collisions and supervise those bad behaviors.
Musk: We hope that over time, Optimus Optimus can become a robot in science fiction movies, just like Star Trek: The Next Generation, just like Data (the humanoid robot character in the Star Trek movie series).
We can program the robot to make it less like a robot and make it more friendly. It can learn to imitate humans and act very naturally. With the general advancement of AI, we can add this feature to robots.
It obviously should be able to execute simple instructions, even intuitively knowing what you want. You can give it high-level instructions, which can break the instructions into a series of actions and execute them.
Q4: Revolving around Optimus Optimus, you think you can achieve several orders of magnitude improvements and economic output, which is really exciting. When Tesla was founded, its mission was to accelerate renewable energy or sustainable transportation.For Optimus Optimus, do you think its mission is still in line with Tesla’s mission statement, or do you need to update the mission to “accelerate the emergence of infinite abundance or the emergence of an unlimited economy”?
Musk: Strictly speaking, Optimus Optimus is not directly consistent with accelerating sustainable energy development.
But it can do the work with a higher efficiency than people, I think it does contribute to sustainable energy development. I think with Optimus Optimus coming, Tesla's mission has indeed expanded to make the future a better place.
Look at Optimus Optimus, I don’t know what you guys think, but I’m very excited to see what Optimus Optimus looks like in the future.
You can judge that for any technology, you want to see what it looks like in 1 year, 2 years, 3 years, 5 years, or 10 years? I want to say, you definitely want to see what Optimus Optimus will look like in the future.
Many other technologies have entered a period of stagnation. I don't want to name it here, but I think Optimus Optimus will become very magical and shocking in 5 or 10 years. I really want to see this happen, and I hope you too.
Q5:I want to know, do you have plans to expand robot dialogue capabilities? The second question is, what is the ultimate goal of Optimus Optimus?
Musk: Optimus Optimus will definitely have dialogue ability. You can talk to it, and the dialogue will feel natural.
From the perspective of the ultimate goal, I don't know, I think it will continue to evolve. I'm not sure what the ending will be, but it must be fun.
We must always be careful and cannot walk the Terminator path. I thought maybe we should start with a video of the Terminator stomping the skull, but people might be too serious about it.
The classic scene in Terminator 2 in which the Terminator stomped the skull to appear
We do hope Optimus is safe, we design some safeguards, you can stop the robot from working through a local read-only memory that cannot be updated through the network. I think it's important, frankly, necessary. It's like a local stop button, and it can't be changed through remote control or something.
But it will definitely be fun and not boring.
Q6: You are displaying very attractive products around Dojo and its applications. I think what the future of the Dojo platform is? Will you provide infrastructure and services like Amazon AWS? Or will they sell chips like Nvidia? What is Dojo's future plan? I see that you use 7nm technology and the development cost can easily exceed US$10 million. What is your business model?
Musk: Dojo is a very large computer that uses a lot of electricity and requires a large amount of cooling devices. I think it might be more reasonable to have Dojo operate in Amazon AWS in the future, rather than selling machines. The most effective way to operate Dojo is to make it a service that can be used online. With it, you can train models faster and more money.
When the world transitions to software 2.0, software 2.0 will use a large number of neural networks for training. This makes sense over time because there will be more neural networks and people will want to use the fastest and cheaper neural network training system. I think there will be many opportunities in this direction.
Q7: What do you think of humanoid robots being able to understand emotions and art and being able to contribute to creativity?
Musk: I think you have seen that robots can at least generate very interesting art, such as DALL·E and DALL·E2 (DALL·E and DALL·E2 are artificial intelligence systems developed by OpenAI that can automatically create real pictures and art through natural language description).
I think we will start to see that AI can even create movies, have coherent, interesting movies, and be able to tell jokes.
Many companies outside Tesla have amazing development speed, and we are heading towards a very interesting future.
Ashok: Optimus Prime Optimus robots can create physical art, not just digital art. You can ask it to do some dance moves in words or voice, and it can create these moves in the future. It's more like physical art than digital art.
Musk: Yes, computers can definitely make physical art.
Ashok: is like dancing, playing football, or something, it needs to be more flexible. As time goes by, it will certainly be done.
Q8: Regarding the introduction of Tesla's autonomous driving system, I noticed that the model you are using is deeply inspired by the language model. I wonder what the history of making this choice is and how much improvement it brings? I previously thought that using language models in lane conversion is a very interesting choice.
John: We transitioned to the language model for two reasons.
first, it allows us to predict lane to a degree that other methods cannot do. As Ashok mentioned before, when we predict lanes in a dense three-dimensional way, you can only model certain types of lanes, but we want to get the crisscrossing road connections at intersections. This is impossible without turning this task into a graph prediction. If you want to complete the task in intensive segmentation, this cannot be successful.
Second, lane prediction is also a multimodal problem. Sometimes you don’t have enough visual information to accurately know the situation on the other side of the intersection, so you need a method that can summarize and generate coherent predictions. You don't want to predict two or three lanes at the same time, you want to predict only one, and generate models, such as these language models, can do this.
Q9: For neural networks, how do you perform software unit testing? Do you have a large number of use cases, thousands of use cases that you must pass these use cases tests before you can release as a product after training the neural network? What is the strategy for your software unit testing?
Ashok:We define a series of tests, starting with unit tests against the software itself, and for the neural network model, we define the test set.
If you only have a large test set, we find that this is not enough. We need complex, test sets for different modes, and then we sort them out and expand this set as the product is used.
Over the years, we have sorted out the once failed use cases, in hundreds of thousands. For any new model, we need to test these failure history and continue to add use cases to this test set.
Above this, we also have shadow modes, we quietly publish these models to the car, and we retract data on when they failed or succeeded.
We also have an extensive testing process. Before pushing it to customers, it must go through 9 layers of filtering. We have a good infrastructure to make all this efficient.
Musk: I am one of the testers and I conduct actual car testing. I've been testing the latest internal beta version in the car to see if it will crash.
Q10:I see these large models, when extending data and model parameters, they can actually make inferences. Do you think that in essence, the basic model needs to be expanded with data and scale, so that at least the "teacher model" can be obtained, which may solve all problems, and then a series of "student models" can be extracted. Do you think of the basic model this way?
Ashok: This is very similar to our automatic annotation model. We don't just run models on cars, we train completely offline, very large models that can't run on cars in real time. We just run these models offline on the server, generate very good annotated data, and then use this data to train the online network. This is the refinement form of these "teacher models" and "student models".
As for the basic model, we are building a very, very large data set with a capacity of up to PB level. We see that when we have these large datasets, some of these tasks run very well, such as the motion parameters I mentioned, input video, output motion parameters of all objects, etc.
People once thought that we could not use the camera to complete the detection, detect depth, speed, acceleration, etc. Imagine how accurate the predictions must be to make these higher order derivatives accurate. It all comes from these large datasets and large models. We regard the basic model as a way to express geometric and motion parameters.
John: Basically, as long as we train based on a large data set, we will see that the model performance will be greatly improved. As long as we initialize our network with a certain pre-training step of some other auxiliary task, we can see improvements. Self-supervised or supervised large data sets are very helpful.
Q11:Eron said Tesla may be interested in establishing a general artificial intelligence AGI system. Given the possible transformational impact of such technologies, investing in security technologies for general AI AGI seems to be a cautious move. I know that Tesla has done a lot of technically narrow AI security research. I want to know whether Tesla intends or tries to lay out the field of general artificial intelligence AGI security technology?
Musk: If we say that we will make a significant contribution to general artificial intelligence AGI, we will definitely invest in security.
I attach great importance to AI security. I think there should be AI regulatory agencies at the government level, just like any that affects public safety has regulatory agencies. We have regulatory agencies for aircraft, cars, food, and drugs because they affect public safety, and AI can also affect public safety.
I think the government hasn't really understood this yet, but I think there should be referees to ensure or try to ensure public safety of general AI AGI.
You can think about it, what are the necessary elements to create a universal artificial intelligence AGI? Access to datasets is extremely important. If you have a large number of cars and humanoid robots that process petabytes video and audio data from the real world just like humans, this is probably the largest data set.
In addition to this, you can obviously scan the Internet incrementally, but the Internet cannot do it, with millions, or even hundreds of millions of cameras, as well as audio and other sensors in the real world.
I think we may have the most data and may have the greatest training computing power. Therefore we may contribute to general AI AGI.
Q12: We did not talk about the electric truck Semi. I want to know, from a perception point of view, what changes are you considering Semi? Compared with cars, the demand is obviously very different. If you don't think so, what's the reason?
Tesla electric truck Semi appeared at the AI day site, but was not specifically mentioned
Musk: I think, no matter what vehicle you drive, what are you needed? What is needed is that biological neural networks and eyes are basically cameras. Your main sensor is two cameras installed on the slow gimbal. The very slow gimbal is your head.
If a biological neural network plus two cameras on a slow gimbal can drive a Semi truck, then if you have 8 360-degree cameras looking around the vision running at higher frame rates and faster reaction speeds, I think it's obvious that you should be able to drive Semi or any vehicle better than humans.
Q13: Assuming Optimus Optimus will be used in different scenarios and evolve at different speeds for these scenarios. Is it possible to independently develop and deploy different hardware and software components for it and deploy them on Optimus Optimus? In this way, will Optimus Optimus develop faster in the future?
Musk: We don’t understand. Unfortunately, our neural networks fail to understand this problem.
Q14:I want to ask the autonomous driving system, when do you plan to promote the FSD beta version to countries outside the United States and Canada? What do you think is the biggest bottleneck or obstacle in the current technology stack of autonomous driving systems? How will you plan to solve the problem so that autonomous driving systems can surpass human driving in terms of performance, safety assurance, and confidence in human use? I remember you also mentioned that you planned to merge highways and cities into a unified technology stack in FSD V11 and make some architectural improvements. Can you explain it in detail?
Musk: From a technical perspective, the FSD beta version should be launched globally by the end of 2022, but for many countries, we need to obtain approval from regulatory authorities. To some extent, we are subject to regulatory approval restrictions in other countries. I think from a technical point of view, the beta version can be prepared to be launched worldwide by the end of 2022.
We expect significant improvements in October 2022, which will be particularly good at evaluating fast moving traffic vehicles, as well as other improvements.
John: In the past mass-produced version of the autonomous driving system and the FSD beta version, and these differences have become smaller and smaller as time goes by. I remember a few months ago, all FSD beta and production versions of autonomous driving systems used the same pure visual object detection technology stack.
There are still some differences between the two at present, the most important thing is how we predict lane. We upgrade the lane model, as I mentioned in my speech, so that it can handle more complex geometric shapes. In the mass-produced version of the autonomous driving system, we still use a simpler lane model.
We are expanding the current FSD Beta model to handle all highway scenarios.
Musk: I drive a Tesla car and use the FSD Beta version. In fact, it has a unified technology stack. It uses the FSD technology stack on both city streets and highways. For me, it performs very well.
We need to verify it in various weather conditions, such as heavy rain, snow, and dust, to ensure that it can perform better than the mass-produced version technology stack in various environments.
We are very close to this goal, and we will definitely be ready by the end of 2022, and it may also be November 2022.
Paul: Based on my personal driving experience, the highway FSD technology stack has far exceeded the mass-produced version technology stack. We hope that by the end of 2022, the parking technology stack will also be included in the FSD technology stack. Basically, before the end of 2022, if you sit in a car in the parking lot, the car will drive to a parking space at the end of the parking lot.
Musk: The basic indicator that needs to be optimized is how much mileage the vehicle can travel before necessary intervention. Significantly increasing the mileage of a vehicle that is fully autonomous before intervention is crucial to safety. This is the basic indicator of our weekly measurements and we have made huge improvements in this regard.
Q15: I am very curious, if you go back to your 20s, what do you hope you will know at that time, and what advice would you give to your younger self?
Musk: I'm trying to think of something useful to say...yes, joining Tesla is one of the things.
I want to try to get in touch with as many great people as possible. I read a lot of books, that's what I do.
I think it doesn't necessarily need to be too volume, there are some benefits to doing this. For me at the age of 20, it might be a good idea to enjoy the moment more and stop occasionally to smell the fragrance of flowers.
For example, when we were developing the Falcon 1 rocket, we developed the rocket on this beautiful island of Kwajalin Atoll (the largest island in the Marshall Islands in the Western Pacific). I didn't even drink a single drink on the beach during that whole time. I would say I should be able to have a drink on the beach and that wouldn't be a problem.
Falcon 1, which is ready for test launch in the temporary launch site of Kwajalin Atoll,
Q16: You use Optimus Optimus to make all robot industry practitioners excited. It feels a lot like the autonomous driving technology 10 years ago, but autonomous driving has proven to be much more difficult than it seemed 10 years ago. Is there anything we didn’t understand 10 years ago but now we can make the robot come sooner?
Musk: In my opinion, general artificial intelligence AGI is developing rapidly, and there is no shortage of important progress almost every week.
At present, AI seems to be able to win in almost all rules-based games, create impressive art, participate in very complex dialogues, write articles, etc. These are constantly improving, and there are still so many talented people studying AI, and the hardware is getting better and better.
is not only our work at Tesla, but AI is on a strong exponential development curve, and it is obvious that we will benefit from it.
Tesla happens to be very good at drives, motors, gearboxes, controllers, power electronics, batteries, and sensors.
As I said, the biggest difference between a robot on four wheels and a robot with hands and feet is to correctly solve the driver problem. This is a question about drivers and sensors, about how to control these drivers and sensors. You have to have all the elements you need to produce a competitive robot, and we are doing that.
Q17: Tesla and you are taking humanity to the next level. You said Optimus Optimus will be used in the next Tesla factory, and my question is, will the new Tesla factory be completely managed by Optimus Optimus? When can people order humanoid robots?
Musk: Yes, I think when starting, we will have Optimus Optimus perform very simple tasks in the factory, such as loading parts. As you can see in the video, transport parts from one place to another, or load parts into our more traditional body welding robot units.
At the beginning, we will try to study how to make it work, and then gradually expand its scope of use.I think Optimus Optimus usage scenarios will grow exponentially and will be very fast.
As for when people can order, I don't know, I don't think it's far away. I think you mean when people can receive robots. I don't know, I will say that it may be 3 years, no more than 5 years, within 3~5 years you may receive Optimus Optimus.
Q18: I think the best way to promote the progress of general artificial intelligence AGI is to let as many smart people as possible participate in the world. Compared with robot companies, considering Tesla's scale and resources and considering the current research status of humanoid robots, isn't it very reasonable for Tesla to open source some simulation hardware? I think Tesla can become the dominant platform provider and the Android or iOS system for the entire humanoid robot research field, rather than just allowing Tesla researchers or their own factory to develop Optimus Optimus, which can open up Optimus Optimus and allow the whole world to explore the research of humanoid robots.
Musk: I think we have to be careful that Optimus Optimus is likely to be used in a bad way, which can happen. You can issue instructions to Optimus, but these instructions are restricted by robot regulations that you cannot violate and cannot cause harm to others. I think Optimus Optimus may bring some security discussion.
Q19: What is the current and ideal controller bandwidth for Optimus Optimus? This event is a powerful advertising for the depth and breadth of the company. What is unique about Tesla that can do this?
David:For bandwidth issues, you must understand or figure out what you want to accomplish the task. If you accept a frequency conversion task, what do you want your limb to do? That's where your bandwidth comes from. This is not a specific number you can directly say, you need to understand your use case, and that is where bandwidth comes from.
Musk: Regarding the bandwidth issue, I think we may eventually increase bandwidth, and the effect is equivalent to improving robot flexibility and reaction time. You can save the robot state, not at 1 Hz speed, but you don't need to increase to 100 Hz level. I don't know what it is, maybe 10 Hz or 25 Hz.
Over time, I think the bandwidth will increase significantly, or it is equivalent to improving flexibility and reducing latency. We hope to minimize latency and maximize flexibility over time.
We are currently a considerable company, and we have many professional and technical developments in many fields to develop to develop electric vehicles and autonomous driving technologies. Basically, Tesla is a combination of several startups, and so far they are almost all successful, so we must have done something right.
I believe that one of the core responsibilities of management companies is to provide an environment for great engineers to grow fully. I think in many companies, maybe most companies, if someone is a really talented engineer, their talent is suppressed in many companies.
In some companies, the way talent is suppressed may not look that bad, because it looks comfortable to work, your salary is high, but the output you need to achieve is so low, which is like a "sweet trap".
In Silicon Valley, there are some "sweet traps". They don't look like bad choices for engineers, but you have to say, a good engineer joins, what output they achieve? The output of these engineering talents seems to be low, even if they seem to be having a happy life. That's why I say there are some "sweet trap" companies in Silicon Valley.
Tesla is not a "sweet trap". We have high requirements. You need to complete a lot of work, and those jobs are cool and not simple.But if you are a super talented engineer, your talents will be more fully utilized than elsewhere. The same is true for
SpaceX.
Q20: The first question, I have been following your progress in the past few years. Today you have made some changes in lane inspection. You said that you used instant semantic segmentation before, but now in order to create lane information, you have established a transfer model. What are the common challenges you still face? For example, as curious engineers, we as researchers can start studying these problems.
The second question, I am very interested in the data engine. You show a case where the car stops driving. How did you find very similar examples from the data you have? It would be great if you could introduce the data engine.
Phil: I will answer the first question first. Take "occupying the network" as an example. You see content in your speech, which did not exist 1 year ago. We only spent 1 year to release more than 12 "occupying network" models.
is a basic model to express the entire physical world under all places and all weather conditions, which is very challenging.
Just over a year ago, we only drove in the two-dimensional world. We used the same static edge to represent the wall and curb, which is obviously not ideal. There is a big difference between the wall and the curb, and when you drive, you make different choices.
After we realized that we had to rethink the whole problem and consider how to solve it. This is an example of the challenges we have to overcome in the past 1 year.
Kate: Back to your second question, how do we actually get tricky examples of parking vehicles, there are several ways to solve it, give two examples.
first, we can trigger divergence signals. The bit that indicates parking will flash between parking and driving states, which will trigger data return.
Second, we can better utilize the logic of the shadow pattern. If the customer ignores a car but we think it should stop because of this, we will also pass this data back.
These are different trigger logics, allowing us to return and get this data.
Q21: There are many companies that are paying attention to the general artificial intelligence AGI problem, which is a very difficult problem. One of the reasons is that this problem itself is difficult to define. Different companies have different definitions, and their focus angles are also different. How does Tesla define general artificial intelligence AGI problems? What are your specific concerns?
Musk: We are not actually paying special attention to the general artificial intelligence AGI problem. I'm just saying that general AI AGI seems likely to be an emerging attribute of what we're doing.
We are creating all these self-driving cars and humanoid robots that are in a huge data stream, data flowing in, processed, which is the largest real-world dataset to date.
These data, you cannot just search the Internet, you must go into the outside world, interact with people, and interact with the road. The earth is a huge place, with chaotic and complicated reality.
I think this seems to be an emerging property. If you have tens of millions or hundreds of millions of self-driving cars, or even a considerable number of humanoid robots, maybe there are more humanoid robots, then this is the largest number of data sets.
These videos are processed and it is very likely that cars will definitely be much better than human drivers, and humanoid robots will likely become increasingly difficult to distinguish from humans.
As I said, you have the emerging attributes of general artificial intelligence AGI.It can be said that all human beings are also super intelligent as a whole, especially when we increase the data interaction rate between humans.
Similar things seem to have happened a long time ago, and the Internet is like the nervous system acquired by humans. Suddenly, by connecting to the Internet, any part of humanity can master all knowledge, almost all knowledge, at least a large part of it.
Before, we exchanged information through individual penetration. In order to transfer data, you have to write a letter, someone has to be responsible for handing the letter to another person, and there is a lot of things happening in the middle. Think about it, it's really slow.
Even if you are in the National Library, you still cannot obtain all the information in the world, and you will definitely not be able to search for this information. And it is obvious that not many people will be in the National Library.
I mean, one of the great equalization elements, the Internet is the largest equalizer in history in terms of obtaining information or knowledge. I think any student studying history would agree with this.
Recently someone asked me, which historical period I like most? My answer is now. Now is the most interesting time in history, and I have read a lot of history books to keep us doing our best to keep this.
Back to the previous question, my answer is, regarding Tesla's autonomous driving system, over time, neural networks gradually digest more and more parts of the software.
Of course, in extreme cases, you can simply get the video you see in the car and compare it with the steering input of the steering wheel and pedal. These are very simple inputs. In principle, you can train directly between the two because that is what humans are doing with biological neural networks.
training video shows the steering wheel and pedal moving, and there is no other software in the middle. We have not reached this level yet, but we are gradually moving in this direction.
Q22: So many people use FSD, in terms of performance statistics, how do you evaluate the company's risk tolerance? Do you think there is a need for greater transparency, or from third-party regulation, to determine what is good enough and to define performance thresholds based on such a large amount of miles?
Musk: Tesla's primary design requirement is safety, which is the case for all aspects.
In terms of automotive machinery safety, we are among all the cars tested by the government, and in terms of passive machinery safety, the probability of injury is the lowest. We also received the highest rating in terms of active safety.
In the future, it will reach this level: the security is surprisingly high, even much better than humans.
Regarding the autonomous driving system, we have indeed published statistical data in a broad sense. According to the mileage statistics, Tesla cars do not turn on the autonomous driving system, and compare the situation of turning on the FSD beta version. We have seen that along the way, the stable improvements of the autonomous driving system have been made.
Sometimes there is such a contradiction: Should we wait until the car is three times safer than humans before deploying this technology?
But I think this is actually a wrong idea in morality. As long as you believe that autonomous driving can reduce harm and death, you have a moral obligation to deploy it, even if you will be prosecuted and blamed by many people.
Because those who have been saved by you do not know that their lives have been saved. Those who occasionally die or get injured, they certainly think that, in any case, it is the fault of the autonomous driving system.
This is why you have to look at the total mileage statistics, how many accidents occur, how many accidents are serious, and how many deaths are caused.
We have more than 3 million cars on the road, and the total mileage is very high every day. It won't be perfect, but importantly, it's obviously safer than not deploying.
Q23: I don’t work in hardware. Maybe the hardware team and you all can give me some inspiration. Why does Optimus Optimus design need to have symmetry? Because we humans have left and right hands, right? We use certain muscle frequency, higher than others, and wear and tear over time. So you may see that some joint failures, or some actuator failures occur more frequently.
As time goes by, I understand that it is still very early. As humans, we have a lot of fantasy and fiction about the superhuman abilities. All of us don’t want to go there directly. We want to stretch our arms and want all these fantasy designs. With everything else in mind, maybe you can take advantage of all these aspects to create more interesting robots in terms of battery and computing strength. I hope you can explore these directions.
Musk: I think it would be cool to turn Detective Gajet (the protagonist of the American cartoon "Detective Gajet") into real, that would be great.
At present we just want to produce a humanoid robot with basic work. Our goal is to produce useful humanoid robots at the fastest speed. I think this will put us on reality and make sure what we do is useful. One of the most difficult things about
is to be useful and then achieve the total value of high efficiency under the curve. For example, how much help you provide to each person on average, multiply by how many people you help in total, and get the total utility value.
strives to release useful products that people like to a large number of people. It is extremely difficult and the degree of difficulty is unimaginable. That's why I say that a company with released products is a huge difference from a company without released products, which is simply a world of difference.
Even if you release a product, can the output value exceed the input cost? This is also extremely difficult, especially for hardware.
I think it would be cool to do something creative over time, such as having 8 arms, publishing different versions. Maybe some hardware companies can add some features to Optimus Optimus, say, maybe a power interface, or something like that. Or you can add attachments to Optimus Optimus, just like adding attachments to your phone.
There may be a lot of cool things to do over time. There may be an ecosystem of many small companies, or even large companies, to produce accessories for Optimus Optimus.
Conclusion
Musk: I want to thank the team for their hard work, you all are amazing. Thank you for coming, and thank you everyone online for your attention.
I think this will be a great video and you can also fast forward to the part you think is the most interesting. But we strive to provide you with a lot of details so that you can watch videos in your spare time. You can focus on the parts you find interesting and skip the others.
Thank you everyone, we will work hard to hold such events every year, we may even do podcasts once a month, but it is great to get everyone involved in this process and show you the cool things that are happening.
"Long-termism" column is updated every Saturday and long holiday, divided into the following series:
Macro theory: politicians, business leaders, etc. in major countries around the world, etc.
Social theory: Bill Gates and other
Social theory: Bill Gates and other
Growth theory: Rockefeller, Carnegie, etc.
Science theory: Nobel Prize winners, Tencent Science WE Conference, etc.
Science theory: Musk, Bezos, Larry Page/Sergey Brin, Zuckerberg, Huang Renxun, Vitalik Buterin, Brian Armstrong, Jack Dorsey, Son Masayoshi, Huawei, Ma Huateng, Zhang Xiaolong, Zhang Yiming, Wang Xing and others
investment statement: Buffett, Munger, Baillie Gifford, Howard Max, Peter Till, Mark Anderson, Katherine Wood, etc.
Management theory: Ren Zhengfei, Kazuo Inamori, etc.
Energy theory: Zeng Yuqun, etc.
l3 Auto says: Li Xiang, He Xiaopeng, Wang Chuanfu, Wei Jianjun, Li Shufu, etc.
Intelligent says: DeepMind, OpenAI, etc.
Meta Universe says: Meta/Facebook, Apple, Microsoft, Nvidia, Disney, Tencent, ByteDance, Epic Games, Roblox, Bilibili/B station, etc.
Interstellar: China National Space Administration, NASA, International Aerospace Conferences over the years, SpaceX, Starlink, Blue Origin, Virgin Galaxy, etc.
Consumption: Amazon, Walmart, Alibaba, JD.com, Pinduoduo, Meituan, Oriental Selection, etc.
Each series focuses on top global experts and industry leaders in various fields, collects and organizes first-hand information such as shareholder letters, public speeches/communications, media interviews, etc., learn classics together, absorbs ideological nourishment, be friends of time, and be long-termists. North American time on September 30, 2022 (Beijing time on October 1, 2022), Tesla held the 2022 AI Day (AI DAY) event at the headquarters in Alto, Palo, California, USA. Tesla CEO Musk led the Tesla team to showcase the prototype of the humanoid robot Optimus, and introduce the latest progress of the fully autonomous driving system FSD, supercomputing platform Dojo, etc.
This issue of "long-termism", choose Tesla's 2022 AI daily record, press conference minutes, on-site Q&A minutes, rubble village man, smart things, car things, electric planet News, Wall Street News released, Liuhe Commercial Research and Selection and Compilation, Share with everyone, enjoy!
Text:
Full text 23,848 words
expected to read for 48 minutes
Tesla 2022 AI Day Record
Word Count: 7,202 Words
Source: Zhidongxi, Car Games, Electric Planet News, Wall Street News
North American time On September 30, 2022, Tesla held an AI Day event. Unlike the previous Musk's lecture mode, on this AI Day, Musk briefly introduced the humanoid robot Optimus that highlighted it, followed by the heads of each business to give a specific explanation. The
activity began. The humanoid robot Optimus Optimus prototype, which had no decorations around, slowly walked onto the stage and completed walking, turning, waving and other actions. Tesla also plays a video of Optimus Optimus, showing that Optimus Optimus can complete transporting items, watering flowers and other activities. According to Musk's vision, Optimus Optimus can be used in the future for homes, cooking, mowing lawns, taking care of the elderly, and even becoming a human partner or companion.
Optimus Prime Optimus appears
Tesla 2021 AI Day, Optimus Prime Optimus is just a concept. In the past year, Tesla has successfully launched a prototype that can walk and complete multiple actions. In order to enable Optimus Optimus to be released smoothly, Musk postponed Tesla's 2022 AI Day for more than 40 days from the original August 19, 2022.
Musk attaches great importance to Optimus Optimus in the future, saying that in the future, Optimus Optimus will be more important than cars and autonomous driving. In the future, Tesla will produce millions of robots, and the price will be cheaper than Tesla car , which may be less than 20,000 US dollars, and consumers will be able to buy it in the next 3 to 5 years. In terms of autonomous driving, Tesla has been improving its technology and products. Currently, the FSD Beta version of the fully autonomous driving system has been used by 160,000 users, and it is expected to have global promotion capabilities by the end of 2022.
Tesla's self-developed supercomputing platform Dojo. Currently, this product is helping Tesla achieve more achievements in autonomous driving, robots, etc.
At the entire press conference, several Tesla engineers introduced Optimus Optimus, FSD, Dojo, etc. in detail software and hardware. When introducing each product, they talked about hoping that more AI, hardware and other talents would join Tesla. Musk also said that the purpose of holding this event is to attract more AI talents to join Tesla and make better products.
Optimus Prime Optimus prototype was unveiled. The expected price is no more than US$20,000
The press conference began. Musk showed the Tesla humanoid robot Optimus Prime Optimus prototype. From the appearance, the robot that debuted first was indeed very prototype, with wires exposed all over the body without any decoration. Optimus Optimus can independently complete walking, wave to users, and even dance.
Optimus Optimus Dancing
Musk said that Optimus Optimus can do more things, but due to the stage, it can only show these. Tesla plays Optimus Optimus video live. In addition to walking around, Optimus Optimus can also complete transportation of items and watering flowers.
Optimus Optimus transports things
In the factory, Optimus Optimus can take a long strip of objects out of the workbench and then neatly discharge them into a box containing the same objects. Judging from the renderings from Optimus's own perspective, it can distinguish different objects in the real world by colors, such as seeing the handheld long strips as purple, the workbench as yellow, etc.
Optimus Optimus transports goods
Tesla then displays a real version of Optimus Optimus that is closer to real people. Its appearance is similar to the 2021 AI Sun Display model, with a human appearance and higher degree of freedom. The real-time Optimus version of Optimus can provide more services, such as free movement of fingers, operating more tools, holding some tools with your right hand, and even doing some repetitive work in the factory. Unlike the initial version, the real version of Optimus Optimus did not walk around the site, but was carried by staff throughout the process.
Musk said that the Tesla robot team has worked almost 7 days a week and worked more than ten hours a day in the past year, and finally successfully displayed the Optimus Prime Optimus prototype. It took only more than 6 months from the launch of the first-generation R&D platform in February 2022 to the launch of the prototype. The Tesla team has done a lot of work, and Optimus Optimus is still in its early stages and can do it better in the future.
real machine version Optimus Prime Optimus
Musk said that Optimus Prime Optimus robot project represents the mission expansion of Tesla. Optimus Optimus is designed based on the human body and will have the ability to communicate and hope that its behavior will be as close as possible to the human behavior. The future full-body version of Optimus Optimus has a weight of 73kg. It uses 100W of electricity when sitting in a meditation, 500W of electricity when walking quickly, and has more than 200 degrees of freedom in the whole body and 27 degrees of freedom in the hands.
Musk predicts that Optimus Optimus can be produced at low cost, with a future output of millions of units, with a cost of less than US$20,000. Compared with similar robot products on the market, the cost will be significantly reduced.
Musk said that robots can reduce labor costs and make social development more motivated. There will be no poverty in the future. Humans can freely choose the type of work and no longer have to do physical work. They can participate more in mental work. I hope that robots can bring more help to humans more safely. Self-driving cars have a huge impact on the world, increasing transportation productivity by at least half an order of magnitude. In the future, robots may have several orders of magnitude improvements.
Tesla's first generation robot has been repeatedly verified from concept, design, analysis, establishment, and optimization. For this new species, having lower cost and higher work efficiency is the key to verifying whether the product can truly be launched into the market.
From the more detailed functions of Optimus Optimus watering flowers and moving things, Tesla insists on constantly evolving towards anthropomorphic robots, from hand movements, gait adjustments, control systems, etc., relying on Tesla's strong technical accumulation in the automotive field, infrastructure and supply chain capabilities, plus Musk's ambitions and strong action power in the field of humanoid robots, bringing unlimited possibilities to the robot industry. After Musk briefly introduced it, the Tesla robot team introduced Optimus Optimus. Starting from Tesla's 2021 AI Day, Optimus has evolved three times so far, and finally achieved results.
Optimus Prime Optimus design is based on the vehicle design process accumulated by Tesla before. Tesla technicians say that cars are wheeled robots, and Optimus Optimus just stands up the car to some extent.
From the Optimus Optimus concept diagram, the entire robot contains 28 electric drives (orange) and battery and control module (blue).
Humans can eat a small amount of food to maintain energy. In order to reduce the energy consumption of robots, Tesla minimizes the energy consumption of humanoid robots when they are idle. Just press the switch to adjust it to be in a low battery or normal working state.
Optimus Prime Optimus power system is integrated into the upper body of the robot, and contains a battery pack with a capacity of 2.3kWh, with a working voltage of 52V, and can run for a whole day after charging. What’s unique about this battery pack is that it integrates all batteries, sensors, etc. using automobiles and energy products. This design is a reference to Tesla's automotive design plan, hoping to reduce more wiring harnesses and focus distribution and calculations on the center of the torso.
Optimus Optimus uses a single Tesla self-developed SoC chip, which supports LTE 4G connection, which is different from the dual-chip autonomous driving solution that Tesla uses in cars. Unlike cars, Optimus Optimus needs to process visual data to respond quickly, which is based on a variety of sensory inputs and communications. Therefore, it is equipped with modules such as radio connection and audio support, which have the characteristics of protecting the robot body and human safety.
Optimus Optimus core components display
In terms of action, Optimus Optimus Optimus absorbs Tesla's automotive power system design experience, and the team analyzes what actions Optimus Optimus needs to take, including walking, going up and downstairs, etc. Then, first analyzes the dynamic data of Optimus Optimus walking, and then analyzes the time, energy consumption, and trajectory required for these movements, and designs the robot joints and actuators based on these data. In terms of security of
, Tesla has made targeted designs to protect robots, R&D personnel optimize their infrastructure. For example, when Optimus Optimus falls, it will not damage the transmission and the arm. After all, the robot repair cost is very high.
Optimus can protect itself when falling
Tesla R&D personnel use the same underlying technology as Tesla cars, making pressure in all components of Optimus Optimus, making it easier to control its walking and not appear stiff.
Optimus Optimus Optimus Walking Pose Simulation
Taking the knee as an example, Optimus Optimus was designed according to the real human knee structure. The R&D personnel simulated the human knee and the force they were exposed during movement, learning how to use less force on the robot knee, allowing it to achieve better force control, and allowing the relevant structure to be tightly wrapped around the knee.
Optimus Prime Simulates human bone structure
Cars and robots have many similarities in power design, so Tesla's experience in power design can be applied to robots. The car drive unit is to accelerate the car. The robot has 28 joint drives, which is not much different from the car drives. However, the tasks that humanoid robots have to do are more complex and require walking or climbing. Therefore, the R&D personnel use models to generate the robot's connection torque speed trajectory, and then enter its optimization model to run.
Cars and Robot Driver Comparison
Robots need to rotate. According to the torque speed trajectory and joint driver efficiency diagram, the energy consumption generated along the track, the cumulative energy of the execution tasks and running time, can define samples of specific actuators and send them to the cloud. This will reduce the time it takes to produce millions of joint drivers.
However, each joint of the robot is specific, and 28 joints require customized specifications. Therefore, R&D personnel need to reduce customized actuator designs, analyze and upload them to the cloud for common research.
Optimus Optimus has 6 types of actuators, including 3 different specifications of servo , 3 different specifications of traction actuators similar to weight scales, etc. Tested within the limits, its joint drive lifts a half-ton piano, which is also a humanoid robot must have functions.
Optimus Prime Optimus joint driver lifts a half-ton piano
Human hands can move at a speed of 300 degrees per second, and have thousands of tactile sensors . Another focus of Optimus Optimus is the hand. Tesla hopes that Optimus Optimus' hands can be as flexible as humans, able to grasp objects, operate, have sensors for perception, etc.
Tesla R&D team also draws inspiration from biology when designing Optimus Optimus's hands. Through 6 actuators, Optimus Optimus can achieve 11 degrees of freedom movement, hold objects weighing 20 pounds (9.1 kg), operate some equipment, or grab small objects, etc.
The adaptation mechanism of the hand is very complex. Humans can recognize the position of the hand in space, which is also the key to its ability to grab objects. Therefore, Tesla is currently conducting corresponding tests.
Optimus Optimus uses the core sensor as the camera, which is similar to the camera used in Tesla's FSD system. Tesla currently collects a lot of data to train Optimus Optimus. In terms of action, Tesla Automobile also uses neural network "occupies the network" to identify the actionable area. After sensing and analyzing the external environment, the software will plan the travel trajectory, and then plan each footing point according to the trajectory, and finally perform the moving action through the actuator.
Optimus Optimus uses the "occupying network" to identify mobile areas
Tesla R&D personnel demonstrate the growth trajectory of Optimus Optimus Optimus's action capability. In April 2022, the first step was taken; in July 2022, the application of a robot unlocking the pelvis to maintain balance; in August 2022, the arms began to work; in September 2022, Optimus Prime Optimus toes also came in handy.
As the humanoid robot slowly utilizes more joints and trains it, the movement speed is significantly improved. Optimus Optimus is currently walking very slowly, not reaching the 5 mph (about 8 km/h) level claimed when it was first released in 2021.
Optimus Prime Optimus action capability growth trajectory
From the perspective of human walking process, it is important for robots to have a physical self-awareness, such as being able to realize the length of their limbs, how to maintain balance, and how to coordinate the movement of their limbs. It's easy for humans to do this, but it's hard for robots.
Optimus Another focus is to keep it upright and not fall to the ground easily. Through the motion planning and control stack, R&D personnel generate robot kinematics models, and then generate the required paths to the underlying platform, allowing the entire system to refer to its trajectory for training. In the Optimus Prime Optimus simulation system, the lines are planned according to its desired path, and the interconnected trajectories are constantly added. Each footing point is planned according to the trajectory, and then the actuator is allowed to execute it, so as to ensure the balance of the robot during walking.
Optimus Prime Walking Upright
In sports training, the motion planning method is an ideal situation, but in fact it is not realistic to put it in the real world.A few key actions are missing in the process, and placing them into the real world can have an impact on model dynamics, especially systems like bipedal dynamics.
R&D personnel use corresponding sensors and observe in the real world to solve robot control problems, use robot pelvic position, center of gravity position, etc. to track the path of the robot in the studio, build a better robot model and correct robot behavior based on actual conditions.
Tesla hopes to make Optimus more flexible in the future, go further from the prototype function, so that it can improve in all aspects, and have better navigation, action and other capabilities.
To solve the problem of humanoid robots manipulating the real object world while observing, R&D personnel divide this problem into two steps, first generate a natural motion reference system, and then optimize and promote it.
Assuming that someone demonstrates a certain action, the R&D personnel set keyframes to its action through video and map it to the robot. This requires only a demonstration video, and R&D personnel can promote it to the robot's realistic manipulation application. This can solve the problems of where the robot should put his hands when grabbing an object, how to carry and place it.
Optimus Optimus performs crawling actions by simulating real people
FSD is progressing smoothly. It is expected to have global promotion capabilities by the end of 2022
In terms of autonomous driving, the Tesla team first introduced the FSD situation of the fully autonomous driving system. FSD Beta, with 2,000 customers participating in the test in 2021, expanding to 160,000 customers in 2022, achieving 80 times growth. Since 2022, Tesla has trained 75,000 neural network models and launched 35 FSD version updates on this basis. In terms of the technical architecture of autonomous driving, Tesla's approach is to first use an automated data annotation system to automatically label the collected data, then process the data to train the neural network, and then deploy the AI model to the FSD computer, calculate the perception results of the external environment through calculations, and then plan the vehicle's driving route. In terms of technology display of
, Tesla first demonstrated its unprotected left-turning ability. For example, when turning left, there are pedestrians and other vehicles. After taking into account the driving trajectories of different traffic participants, Tesla calculated the most suitable driving trajectory.
Tesla FSD no protection left turn capability
Tesla uses the name interactive search technology. First, start with visual perception, perceive traffic participants, then infer their driving trajectory, then generate several strategies, and finally select the best driving trajectory.
Tesla interactive search technology architecture
It should be noted here that if there are more and more external targets, the amount of calculations will become larger and larger. Tesla perceives the outside world by looking around the camera, generates a 3D environment, and finds the driving areas through the "occupying the network" to identify which obstacles are. When working in
FSD, the camera image is calibrated first, the images are integrated together, the images are formed, the data is extracted, and the data is input into the neural network. Through the corresponding algorithm, spatial features are constructed. After
only generates 3D space, there is no precise location of various objects, and it is still impossible to plan the path. Therefore, Tesla analyzes key features and calculates the location data.
Tesla FSD builds a 3D space based on visual data
Tesla team accumulates a lot of video clips during daily driving, and each video has many frame images.Tesla needs 1.4 billion images to train a neural network, and needs to use 100,000 GPU working hours (1 GPU 1 hour) to work. The computing power is huge and supercomputers and AI accelerators are required. This is also an important reason why Tesla's self-developed supercomputing platform Dojo. Based on Dojo, Tesla can increase the training speed of neural networks by 30%. In terms of predicting the behavior of other traffic participants, Tesla's approach is to first enter the RegNet network ( convolutional neural network developed by Meta's FAIR laboratory), and then the processed data enters the Transformer model ( deep learning model developed by Google ). There may be 1 billion parameters to optimize them together. To achieve maximum computing power and minimize latency.
Tesla cars will generate a large amount of data during operation, and these data also need to be marked. In terms of data annotation, Tesla first tried manual annotation, which was time-consuming and labor-intensive; then considered the supplier cooperation method. Judging from the final results, both timeliness and quality are not very good. Tesla needs very efficient and scalable annotations.
Tesla currently adopts the standard method of human-machine cooperation, including both human and machine annotations. Overall, machine labeling is more efficient, and the machine has a 30-minute workload, which may take a long time for humans, so Tesla is building an automatic labeling system. Through efficient annotation, space-time fragments in the real world are converted into usable data, making FSD more intelligent and efficient. Automatically labeling of data also needs to be sorted out. Tesla did not invest much energy in this area before, and now many engineers are doing this.
Tesla trains the model based on automatic labeling data
simulation system is a very important part of the autonomous driving system, which can improve the vehicle's ability to respond to long-tail scenarios. Tesla has developed a scene generator that can generate a scene in as fast as 5 minutes. Compared with the traditional method, the speed is 1,000 times higher. It can also scan real objects to project onto the screen, simulate signal lights, stop signs, etc., and be as close to the real world as possible. This is of great significance to model training.
Tesla R&D Scene Generator
Currently, Tesla data sets are partly from the information sent back by the fleet and partly from the simulation data, which can make judgments on the scene more convenient. Through the data engine, neural networks can be made more realistic, bring more certainty to FSD, and solve uncertainty in the real world. If a car turns at an intersection, it is necessary to determine whether the parking vehicle is parked or is driving slowly. Just create more networks for evaluation to solve this scenario problem.
Regarding the promotion of Tesla's FSD beta version, Tesla said that by the end of 2022, it will have the ability to promote FSD globally. In addition to North America, Tesla also needs to communicate with local regulatory authorities. In some countries and regions, the regulatory policies for autonomous driving are still lagging behind.
supercomputing platform Dojo is constantly iterating to ensure Tesla's computing power needs
Tesla team has mentioned the supercomputing platform Dojo many times when introducing Optimus Prime Optimus robot and fully autonomous driving FSD.
Tesla 2021 AI Day, the first Tesla's first AI training chip Dojo D1 is displayed for the first time, and the supercomputer system ExaPOD built on this chip can be used to perform AI training tasks and provide support for the huge image processing needs of on-road vehicles.
Tesla currently has a large supercomputing platform based on Nvidia GPU, and a data center that stores 30PB video materials. It is currently developing a supercomputing platform based on Dojo chips.
Tesla uses a set of pictures to show the key nodes of Dojo on the supercomputing platform in the past two years, from the delivery of a customized refrigerant distribution unit CDU, to the installation of the first integrated Dojo cabinet, and then to the 2.2MW unit load test.
Supercomputing platform Dojo Key node
Tesla tries to optimize the scalability of Dojo design and overcome challenges in a fast trial and error manner. Dojo has features such as a single scalable computing plane, global addressing fast memory, unified high bandwidth + low latency, etc.
Tesla technical engineer, especially when talking about the Dojo voltage regulation module. In the past two years, the voltage regulation module has been updated in 14 versions, with high performance, high density, complex integration and other characteristics.
supercomputing platform Dojo voltage regulation module
thermal expansion coefficient CTE is very important. Tesla cooperates with suppliers to provide power solutions to reduce the thermal expansion coefficient CTE of the supercomputing platform Dojo over 50%, making Dojo performance reach 3 times the initial version.
Tesla team demonstrates the use of Dojo to help AI drawing software Stable Diffusion to draw Tesla electric pickup Cybertruck to exercise images on Mars.
Cybertruck on Mars picture
According to reports, only 4 Dojo cabinets can replace 72 GPU racks of 4,000 GPUs. Dojo can reduce the working hours usually take several months to 1 week.
Tesla's self-developed D1 chip also plays an important role in the supercomputing platform Dojo. The D1 chip adopts TSMC's 7nm process process , and 50 billion transistors are distributed over an area of 645mm². The peak computing power reaches 362 TFLOPS (BF16/CFP8 accuracy, BF16 and CFP8 are floating point calculation format standards), and the thermal design power consumption TDP does not exceed 400W.
is based on D1 chip. Tesla launched a system-level solution on wafers. By applying TSMC InFO_SoW packaging technology, 25 D1 chips are used to form a training module Training Tile. The peak computing power of one training module reaches 9 PFLOPS (with BF16/CFP8 accuracy), and includes computing, I/O (input/output), power, liquid cooling and other modules. It adopts centralized power supply and heat dissipation design, and the heat dissipation power reaches 15kW.
supercomputing platform Dojo's training module Training Tile
Dojo system tray System Tray has high-speed connection, dense integration and other characteristics. The peak computing power can reach 54 TFLOPS (with BF16/CFP8 accuracy), and the power consumption is 100+kW.
Dojo interface processor is a PCIe card with high bandwidth memory, using Tesla's own TTP interface. Tesla transmission protocol TTP can bridge to standard Ethernet, and TTPOE can convert standard Ethernet to Z-plane topology , with high Z-plane topology connectivity.
Tesla 2021 AI Day, Dojo development has ushered in a series of milestones, including installing the first Dojo cabinet and conducting 2.2mW load testing. Tesla is currently promoting Dojo R&D at the speed of creating a training module every day.
supercomputing platform Dojo cabinet
Tesla will form a Dojo supercomputer with 60 training modules, with 1,500 D1 chips and more than 530,000 training nodes. In theory, there is no upper limit for Dojo performance expansion and can be expanded infinitely.
In actual applications, Tesla will form 120 training modules to form a supercomputer system ExaPOD, with 3,000 D1 chips and more than 1 million training nodes, with peak computing power exceeding Level E, reaching 1.1 EFLOPS (with BF16/CFP8 accuracy), and also provide 13TB memory and 1.3TB cache capability.
Tesla team said that computing power is the foundation of fully autonomous driving, and high-efficiency chips can better serve fully autonomous driving. What Tesla needs to do is to make Dojo the strongest supercomputer system in AI training.
Tesla does not limit Dojo design. It can provide a very large and efficient system, breaking the physical boundaries of traditional integration from the hardware, and making the hardware more efficient on the compiler. As long as physical conditions allow, Tesla can continuously break through the limit.
Tesla announced that it is expected to deploy the first ExaPOD in Q1 2023. After successful construction, ExaPOD will become one of the most powerful supercomputers in the world.
Tesla said that with the exclusive compiler, DOJO training delay can be achieved at least 1/50 of the same-scale GPU. Tesla's goal is that by mass production in Q1 2023, DOJO can achieve 4.4 times the single-chip training speed of Nvidia A100, and even lower energy consumption and cost.
Tesla 2022 AI Day Press Conference Minutes
Word count: 5,656 words
Source: Ruble Village Man
Tesla 2022 AI Day, after Musk briefly introduced the humanoid robot Optimus Optimus, the heads of Tesla's business gave a specific explanation.
humanoid robot Optimus
Musk: Welcome to Tesla AI Day 2022. We have some very exciting content to show you and I think it will impress you.
For Optimus Optimus Robot, I do want to set expectations. In 2021, it was just a guy wearing a robot suit and now we have made great progress. Compared to 2021, it will be impressive. We will talk about our progress in AI, autonomous driving, and Dojo.
Should we let the robot appear?
Kate: Before that, we had a small tip, which is actually the first time we tried this robot without any fallback support. Cranes, mechanical devices, cables, etc., nothing.
Musk: We will show you videos of robots doing other things.
Milan: We want to show some of the progress made around robots in the past few months, which can move around and dance.
This is just a small start. You can see that the autonomous driving neural network is running as it is. We are just retraining the robot directly on the new platform.
Musk: When you see the rendered view, that is the world that the robot sees. It can identify objects very clearly, such as the ones it should pick up.
Milan: We use the same process as the autonomous driving system to collect data and train neural networks. For robots, we also deploy this way.
This is an example to further demonstrate the robot's upper body function. We do want to improve this part of the functionality in the next few months.
Kate: That is not the only content we want to show today.
Musk: The robot you just saw is called Wasp Bumble C. It is a development version of the robot, which uses a semi-off-the-shelf drive.
In fact, we have gone a step further and our team has done an amazing job. In fact, we have an Optimus Optimus robot that uses drives, battery packs, control systems and everything else that is completely designed and produced by Tesla. It can't walk, but it will be within a few weeks.
We want to show this robot, but in fact, it is quite close to the production standard. We want to show all the things it can do, let's ask for a robot.
You can see Optimus Optimus, the degree of freedom it has is what we hope the first mass production machine can have. That is, the ability to move all fingers independently, the thumb has two degrees of freedom. It has an opposing thumb, a left hand and a right hand, and can operate tools and do useful things.
Our goal is to produce useful humanoid robots as soon as possible.When designing it, we adopt the same principle as designing cars, which is designed for production. This is the only way to produce robots with high output, low cost and high reliability.
Optimus Optimus design goal is to produce millions of robots with extremely strong capabilities and extremely high production. Optimus Optimus is expected to cost much less than a car, probably less than $20,000, which is my guess.
Its potential is unfathomable. You can say, what is economy? The economy is the number of entities that carry out production multiplied by productivity and the population multiplied by per capita output. Once there is no limit on the population, what the economy actually means will be less clear and the economy will be close to infinity.
This means a rich future, a poor future. At that time you could have any product and service you wanted, which was indeed a fundamental transformation of human civilization as we know it.
It is very important that corporate entities that turn this ideal into reality need to accept the reasonable impact brought by the public. I think Tesla's structure is very ideal.
Kate: You have seen several robots today, let's quickly review the timeline.
appears for your robots who perform talents. We will complete the production within 6 months, and carry out software integration and hardware upgrades within 1 month thereafter.
At the same time, we are also designing the next generation of robots, this robot here. This guy, rooted in the vehicle design process, is leveraging all of this experience already.
Repeat, we adopt the basics of vehicle design, from concept to design and analysis, and then build and verify. In this process, we will optimize cost and efficiency, and ultimately these are key indicators for the product to scale.
Inside the torso, we install the battery pack with a capacity of 2.3kWh, which is a perfect configuration for all day use. Next is the brain, which is not on the head, but is very close to it, in the torso, we install the central computer.
Tesla has equipped every car with an FSD system. We want to use the hardware and software of the autonomous driving system to develop a humanoid robot platform. But because its needs are different from appearance factors, we must first make changes.
It requires doing everything the human brain does, including processing visual data, making instant decisions based on multi-sensor input, and communication. To support communication, it is equipped with wireless connection and audio support. It also has hardware-level security features, which is important for protecting robots and people around them.
Nilegen: Can we use our capabilities and methods on the automotive side to influence robots?
Since we have collision software, we can use the same software to make it fall down. The purpose of this is to ensure that even if the robot falls, it will only be damaged on the surface. Of course, it is best not to fall. We hope it dusts and continues to complete the task.
drive can lift a concert grand piano weighing half a ton and 9 feet long. The design of the robot hand is inspired by biology and has 5 fingers, driven by metal tendons, flexible and strong, with the ability to grasp a large range of force, and is also optimized for precise grasping of very thin and small objects.
Milan:We show all this cool content in the video, all of which are done in just a few months. Thanks to our magical progress on autonomous driving systems over the past few years, most of the components can be easily ported to a robotic environment.
You can think of it, this is just turning the robot on the wheel into a robot with long legs, some of which are very similar, others require us to put in more work.
For example, our computer vision neural network is directly transplanted from the autonomous driving system to the robot environment. We are also working hard to find ways to use neural radiation fields to improve these "occupying networks" and obtain a good volume rendering effect for the surrounding environment of the robot. For example, a robot interprets what it needs to interact with.
Another interesting question is, in indoor environments, there is no GPS signal in most cases, how to make it navigate to the destination?
We have been training more neural networks to identify high-frequency features and key points in the robot camera to obtain images, and track across frames and time in the robot navigation environment. We use these points to better estimate robot poses and trajectories in the work environment.
This is a video that demonstrates the code to run motion control in an autonomous driving system simulator and demonstrates the evolution of the robot's walking ability. As can be seen, it walked quite slowly when it started in April 2022. Over the past few months, it has begun to accelerate as more joints are unlocked with more advanced technologies such as arm balance.
Hopefully so far, you have a good understanding of our work in the past few months. We started implementing a usable robot, but it was far from practical. There is still a long and exciting road ahead of us.
I think in the next few weeks, we need to complete the first task, to get Optimus Optimus to at least reach or even surpass the Wasp level, which is the robot prototype you just saw.
We will also start to focus on real usage scenarios in one of our factories, committed to truly solving problems, and thoroughly consolidate all the elements of deploying this product to the real world, including the indoor navigation, elegant comprehensive management, and even providing services, and all the components required for large-scale production.
I don't know what you think, but after seeing what we showcased, I'm sure we can achieve this goal in the next few months or years, turn this product into reality, and change the entire economy.
I thank the entire Optimus Optimus team for working hard over the past few months and I think they work very well. All of this was completed in just 6 to 8 months, thank you very much.
Automatic driving/FSD Beta version
Ashok:
In the past year, we have trained 75,000 neural network models and completed one model training every 8 minutes. We evaluated on a large computer cluster and released 281 of them, which indeed improved the performance of the car.
This innovation speed appears in all aspects of the entire technology stack, including planning software, infrastructure, tools, etc. Everything is developing to a higher level.
We take this intersection scenario as an example to explore how autonomous driving systems plan and make decisions.
We walk from the side path to the intersection and have to make way for all vehicles crossing the road. Just as we were about to enter the intersection, pedestrians on the other side of the intersection decided not to walk the zebra crossing and cross the road. Now we must give way to this pedestrian and to the vehicles coming to the right. We also need to understand the relationship between pedestrians and vehicles on the other side of the intersection. We need to quickly judge the dependence relationship between a large number of objects.
Humans are very good at this. We see a scenario that understands all possible interactions, evaluates the most likely interactions, and usually ultimately chooses a reasonable judgment. But the frame can also be expanded to the objects behind the occlusion.
We use video sources from 8 cameras to generate three-dimensional occupation information of the surrounding world. The blue part here, corresponding to what we call the "visible area", is basically blocked by the first obstruction you see in the scene. We use the model to generate "ghost objects" in the "visible area". If you correctly model the generation area and state transitions of "ghost objects", if you adjust the control reaction as a function of the possibility of existence, you can extract some very good human-like behaviors.
Phil: "occupy the network" receives all 8 of our camera video streams as input, and directly generates a unified occupancy rate in the vector space. For each three-dimensional position around our car, the probability that that position will be occupied is predicted.
Tim: Let's talk about the infrastructure for training. We have watched four or five videos, and the number of video clips I think about and care about is much larger than that.
We just watched the "occupying network" introduced by Phil. This video alone requires 1.4 billion images to train the network you just saw. If you have 100,000 GPUs, you need 1 hour; but if you only have 1 GPU, you need 100,000 hours. This training task takes a length of time, not something you can afford.
We hope to publish it faster, which means that we need parallel processing, we need greater computing power, which means we need supercomputers.
This is why we built 3 supercomputers within the company, including 14,000 GPUs. We use 10,000 of these GPUs for training, and about 4,000 GPUs are used for automatic annotation.
I can keep talking. I just briefly introduced our two internal projects, which are actually just part of the huge project to optimize our internal computing power.
Through the accumulation and integration of all these optimizations, we now train "occupy the network" twice as fast as it is because its efficiency is doubled. If we add more computing power and use parallel computing, we can complete the training in a few hours, not a few days.
John: I am the head of the visual team of the autonomous driving system. Today I want to introduce two topics to you. First, how do we predict the lane; second, how do we predict the future behavior of other objects on the road.
What we obtain through this lane detection network is a series of lane connectivity, which is directly calculated and output by the network. There are no additional steps here, and there is no need to apply intensive predictions to decentralized predictions, which is the direct output of an unfiltered network.
Above I discuss some content about lane detection. I will briefly discuss how to model and predict the future paths of other objects. I want to quickly show two examples. In the video on the right of
, there is a car running a red light and turning in front of us. Our approach to this situation is to make a series of short-term cycles of future trajectory predictions for all objects. We can use these results to predict possible dangers and use braking, steering and other behaviors to avoid collisions.
Overall, the visual technology stack of the autonomous driving system predicts not only the geometric motion parameters of the surrounding world, but also the rich semantics, thereby achieving safe driving similar to humans.
Jaegan: I will talk about automatic annotation. We have several automatic annotation frameworks that support various types of networks. Today I would like to focus on this excellent lane network.
This network is easy to expand as long as we have enough computing power and itinerary data. In this scenario, about 50 trips are automatically marked, some of which are shown here, 50 trips from different vehicles.This is the process of our capture and transforming space-time fragments of the world into network supervision.
David: Take the simulation scene played behind me as an example. It takes an artist 2 weeks to complete the design. This is too slow for us.
I will talk about using Jaegan's automatic benchmark annotation and some brand new tools. We can generate this scene in just 5 minutes and processically and many similar scenes. This speed is amazing, 1,000 times faster than before.
This method is for scale and scale, be prepared. As you can see on the map behind, we can easily generate most urban streets in San Francisco without spending months or even years, just one person working for 2 weeks.
We reviewed that because we generate all fragment data sets through benchmark data, including all the complex situations in the real world, we can combine process-based vision and various changes in traffic conditions to create infinite target data for online learning.
Kate: This data engine framework is suitable for all signals, whether it is three-dimensional multi-camera video, whether the data is manual standard, automatic annotation, or simulated data, whether it is offline model or online model.
Tesla can optimize on a large scale, thanks to the advantages of the fleet, the infrastructure built by our terminal team, and the labeling resources provided for our network. To train all this data, we need a lot of computing power.
supercomputing platform Dojo
Pete: I am often asked why a car company wants to build a supercomputer for training?
raises this question, but it is still a fundamental misunderstanding of the essence of Tesla. Essentially, Tesla is a hard-core technology company.
Yaji:
Since then, the entire team has been working hard and committed to large-scale deployment. Now we have made amazing progress, and we have achieved many milestones throughout the process, and we have encountered many unexpected challenges. It is our philosophy of "quick trial and error" that allows us to break through our own limits.
starts with our custom D1 chip uniform nodes, connect them to our fully integrated training module, and then finally seamlessly connect them across the boundaries of the cabinet to form the supercomputing platform Dojo.
In short, one ExaPOD can accommodate two Dojos, and the overall computing power reaches 1.1 EFLOPS. In computing history, this degree of technology and integration have only appeared a few times.
Rajeef: This operation takes only 5 microseconds on 25 Dojo chips, while the same operation takes 150 microseconds on 24 GPUs, which is an order of magnitude improvement compared to the GPU. How are the two networks performing? We are going to see the results, which are calculated on multi-chip GPU and Dojo systems, but are both normalized to single chip values.
On our automatic labeling network, the previous generation of VRMS software running on the current hardware can surpass the performance of Nvidia A100; on our production hardware, we can run our newer VRMS software to achieve twice the throughput of A100. Our model shows that with some key compiler optimizations, we can achieve more than 3 times the performance of the A100.
We have seen a bigger leap in "occupying the network". Using our production hardware, we have almost achieved a 3x performance improvement, and there is still more room for improvement.
used to take more than a month to train the network, but now it only takes less than a week.We start with hardware design, breaking through the boundaries of traditional integration and serving our vision of a single giant accelerator. We have seen how to build a compiler on this hardware.
Dojo performance is proven through these complex real-world networks. We also know what our first large-scale deployment should be, our high computing strength automatic annotation network.
Today these networks use 4,000 GPUs on 72 GPU racks. With our intensive computing power and high performance, we hope to provide the same computing throughput with just 4 Dojo cabinets. These 4 Dojo cabinets will be part of the first ExaPOD, which we plan to deploy in Q1 2023, more than twice the existing automatic annotation capability of Tesla.
The first ExaPOD is part of the 7 ExaPODs we plan to build in Palo Alto, California, USA, just across the wall. We have 1 of the ExaPOD display cabinet for everyone to watch.
Conclusion
Musk: We really want to show the depth and breadth of Tesla technology, as well as AI, computing hardware, robots, drives, etc.
We strive to change people's perception of Tesla. Many people think that we are just a car company and we only make cool cars. Most people don’t know that Tesla can be said to be a global leader in AI, hardware and software.
We are building arguably the first, possibly the most radical computer architecture since the Cray-1 supercomputer (the fastest supercomputer in the world from 1976 to 1982).
If you are committed to developing the world's most advanced technology and truly influence the world in a positive way, joining Tesla is the right thing to do.
Tesla 2022 AI Day Live Q&A Minutes
Word Count: 10,990 words
Source: Ruli Village
Tesla 2022 AI Day, Musk led the Tesla team to introduce Optimus Prime Optimus prototype, fully autonomous driving system FSD, and supercomputing platform Dojo, and answered 23 sets of questions raised by the audience.
Musk: I hope we will tell enough details and accept the question now.
Q1: Optimus Optimus left a deep impression on me. I wonder why my hands use rope driving method? Why do you choose rope driving methods for your hands? Because the mechanical tendon is not very durable. Also why spring-loaded models?
Mike: This is a good question. When evaluating any kind of drive scheme, whether it is a rope drive system or a connecting rod system, there must be some choice.
The main reason we chose the rope drive system is that first we investigated some synthetic tendons and found that the strength of metal marine cables is much higher. One of the advantages of these cables is that they can reduce energy consumption very well.
We want to produce a large number of hands. When mass production is carried out, many parts and many small connecting rod devices will eventually become a problem. An important reason why rope drive is better than connecting rods is that it can eliminate gaps. The essence of gap elimination is to keep your fingers from lagging when they move. The main benefit of the
spring loading model is that it allows us to actively open our hands. We don't need to use two drivers to drive the fingers to close and open, we have the ability to let the mechanical tendons drive them to close and then the spring is passively elongated.
This point can also be seen in our hands. We have the ability to actively bend our fingers and also have the ability to stretch our fingers.
Musk: Our purpose of designing Optimus Optimus is to realize the robot that plays its greatest role as soon as possible.
There are many ways to solve various problems of humanoid robots. We may not have found the correct answer in all technical solutions. I should say that we are open to the technological solutions you see evolve over time, and they are not static.
But we have to make a choice, we want to choose something that will allow us to achieve production as soon as possible, as I said, it can work as soon as possible. We strive to follow our goals to produce useful robots at the fastest speed and can be mass-produced.
We will conduct internal testing of this robot in the Tesla factory to see how effective it is. We need to complete a closed loop in reality to ensure that the robot is useful.
We are confident that we can use the hands we are currently designed to achieve this goal, but it is certain that the hand design will have the second and third editions. Over time, we may make considerable changes to the robot structure.
Q2: Optimus Prime Optimus robot is indeed impressive, and bipedal robots are indeed difficult. But I noticed that you may lack recognition of the spiritual value of human beings in your plan. I wonder if Optimus Optimus has personality and can be amused by us when folding clothes for us?
Musk: We hope there is a really interesting version of Optimus. Optimus Optimus can be utilitarian, can complete tasks, or play with you like friends and partners.
I believe people will come up with various creative uses of this robot. Once the core intelligence and drive problems are solved, you can put on various vests on the robot and replace the skin of the robot in different ways.
I believe people will come up with various interesting versions of Optimus Optimus.
Q3: I want to know if there is any behavior equivalent to manual intervention for Optimus Optimus. It seems that it is important to mark moments when humans and machines have different judgments. For humanoid robots, this may also be an ideal source of information?
Ashok:I think we will have multiple ways to remotely manipulate the robot and intervene when it goes wrong, especially when we are training the robot.
We hope to design it in some way that if it is going to hit something we can press a button and it will stop without crushing your hand or something, which are intervening data.
We can also learn a lot from our simulation system, which can check collisions and supervise those bad behaviors.
Musk: We hope that over time, Optimus Optimus can become a robot in science fiction movies, just like Star Trek: The Next Generation, just like Data (the humanoid robot character in the Star Trek movie series).
We can program the robot to make it less like a robot and make it more friendly. It can learn to imitate humans and act very naturally. With the general advancement of AI, we can add this feature to robots.
It obviously should be able to execute simple instructions, even intuitively knowing what you want. You can give it high-level instructions, which can break the instructions into a series of actions and execute them.
Q4: Revolving around Optimus Optimus, you think you can achieve several orders of magnitude improvements and economic output, which is really exciting. When Tesla was founded, its mission was to accelerate renewable energy or sustainable transportation.For Optimus Optimus, do you think its mission is still in line with Tesla’s mission statement, or do you need to update the mission to “accelerate the emergence of infinite abundance or the emergence of an unlimited economy”?
Musk: Strictly speaking, Optimus Optimus is not directly consistent with accelerating sustainable energy development.
But it can do the work with a higher efficiency than people, I think it does contribute to sustainable energy development. I think with Optimus Optimus coming, Tesla's mission has indeed expanded to make the future a better place.
Look at Optimus Optimus, I don’t know what you guys think, but I’m very excited to see what Optimus Optimus looks like in the future.
You can judge that for any technology, you want to see what it looks like in 1 year, 2 years, 3 years, 5 years, or 10 years? I want to say, you definitely want to see what Optimus Optimus will look like in the future.
Many other technologies have entered a period of stagnation. I don't want to name it here, but I think Optimus Optimus will become very magical and shocking in 5 or 10 years. I really want to see this happen, and I hope you too.
Q5:I want to know, do you have plans to expand robot dialogue capabilities? The second question is, what is the ultimate goal of Optimus Optimus?
Musk: Optimus Optimus will definitely have dialogue ability. You can talk to it, and the dialogue will feel natural.
From the perspective of the ultimate goal, I don't know, I think it will continue to evolve. I'm not sure what the ending will be, but it must be fun.
We must always be careful and cannot walk the Terminator path. I thought maybe we should start with a video of the Terminator stomping the skull, but people might be too serious about it.
The classic scene in Terminator 2 in which the Terminator stomped the skull to appear
We do hope Optimus is safe, we design some safeguards, you can stop the robot from working through a local read-only memory that cannot be updated through the network. I think it's important, frankly, necessary. It's like a local stop button, and it can't be changed through remote control or something.
But it will definitely be fun and not boring.
Q6: You are displaying very attractive products around Dojo and its applications. I think what the future of the Dojo platform is? Will you provide infrastructure and services like Amazon AWS? Or will they sell chips like Nvidia? What is Dojo's future plan? I see that you use 7nm technology and the development cost can easily exceed US$10 million. What is your business model?
Musk: Dojo is a very large computer that uses a lot of electricity and requires a large amount of cooling devices. I think it might be more reasonable to have Dojo operate in Amazon AWS in the future, rather than selling machines. The most effective way to operate Dojo is to make it a service that can be used online. With it, you can train models faster and more money.
When the world transitions to software 2.0, software 2.0 will use a large number of neural networks for training. This makes sense over time because there will be more neural networks and people will want to use the fastest and cheaper neural network training system. I think there will be many opportunities in this direction.
Q7: What do you think of humanoid robots being able to understand emotions and art and being able to contribute to creativity?
Musk: I think you have seen that robots can at least generate very interesting art, such as DALL·E and DALL·E2 (DALL·E and DALL·E2 are artificial intelligence systems developed by OpenAI that can automatically create real pictures and art through natural language description).
I think we will start to see that AI can even create movies, have coherent, interesting movies, and be able to tell jokes.
Many companies outside Tesla have amazing development speed, and we are heading towards a very interesting future.
Ashok: Optimus Prime Optimus robots can create physical art, not just digital art. You can ask it to do some dance moves in words or voice, and it can create these moves in the future. It's more like physical art than digital art.
Musk: Yes, computers can definitely make physical art.
Ashok: is like dancing, playing football, or something, it needs to be more flexible. As time goes by, it will certainly be done.
Q8: Regarding the introduction of Tesla's autonomous driving system, I noticed that the model you are using is deeply inspired by the language model. I wonder what the history of making this choice is and how much improvement it brings? I previously thought that using language models in lane conversion is a very interesting choice.
John: We transitioned to the language model for two reasons.
first, it allows us to predict lane to a degree that other methods cannot do. As Ashok mentioned before, when we predict lanes in a dense three-dimensional way, you can only model certain types of lanes, but we want to get the crisscrossing road connections at intersections. This is impossible without turning this task into a graph prediction. If you want to complete the task in intensive segmentation, this cannot be successful.
Second, lane prediction is also a multimodal problem. Sometimes you don’t have enough visual information to accurately know the situation on the other side of the intersection, so you need a method that can summarize and generate coherent predictions. You don't want to predict two or three lanes at the same time, you want to predict only one, and generate models, such as these language models, can do this.
Q9: For neural networks, how do you perform software unit testing? Do you have a large number of use cases, thousands of use cases that you must pass these use cases tests before you can release as a product after training the neural network? What is the strategy for your software unit testing?
Ashok:We define a series of tests, starting with unit tests against the software itself, and for the neural network model, we define the test set.
If you only have a large test set, we find that this is not enough. We need complex, test sets for different modes, and then we sort them out and expand this set as the product is used.
Over the years, we have sorted out the once failed use cases, in hundreds of thousands. For any new model, we need to test these failure history and continue to add use cases to this test set.
Above this, we also have shadow modes, we quietly publish these models to the car, and we retract data on when they failed or succeeded.
We also have an extensive testing process. Before pushing it to customers, it must go through 9 layers of filtering. We have a good infrastructure to make all this efficient.
Musk: I am one of the testers and I conduct actual car testing. I've been testing the latest internal beta version in the car to see if it will crash.
Q10:I see these large models, when extending data and model parameters, they can actually make inferences. Do you think that in essence, the basic model needs to be expanded with data and scale, so that at least the "teacher model" can be obtained, which may solve all problems, and then a series of "student models" can be extracted. Do you think of the basic model this way?
Ashok: This is very similar to our automatic annotation model. We don't just run models on cars, we train completely offline, very large models that can't run on cars in real time. We just run these models offline on the server, generate very good annotated data, and then use this data to train the online network. This is the refinement form of these "teacher models" and "student models".
As for the basic model, we are building a very, very large data set with a capacity of up to PB level. We see that when we have these large datasets, some of these tasks run very well, such as the motion parameters I mentioned, input video, output motion parameters of all objects, etc.
People once thought that we could not use the camera to complete the detection, detect depth, speed, acceleration, etc. Imagine how accurate the predictions must be to make these higher order derivatives accurate. It all comes from these large datasets and large models. We regard the basic model as a way to express geometric and motion parameters.
John: Basically, as long as we train based on a large data set, we will see that the model performance will be greatly improved. As long as we initialize our network with a certain pre-training step of some other auxiliary task, we can see improvements. Self-supervised or supervised large data sets are very helpful.
Q11:Eron said Tesla may be interested in establishing a general artificial intelligence AGI system. Given the possible transformational impact of such technologies, investing in security technologies for general AI AGI seems to be a cautious move. I know that Tesla has done a lot of technically narrow AI security research. I want to know whether Tesla intends or tries to lay out the field of general artificial intelligence AGI security technology?
Musk: If we say that we will make a significant contribution to general artificial intelligence AGI, we will definitely invest in security.
I attach great importance to AI security. I think there should be AI regulatory agencies at the government level, just like any that affects public safety has regulatory agencies. We have regulatory agencies for aircraft, cars, food, and drugs because they affect public safety, and AI can also affect public safety.
I think the government hasn't really understood this yet, but I think there should be referees to ensure or try to ensure public safety of general AI AGI.
You can think about it, what are the necessary elements to create a universal artificial intelligence AGI? Access to datasets is extremely important. If you have a large number of cars and humanoid robots that process petabytes video and audio data from the real world just like humans, this is probably the largest data set.
In addition to this, you can obviously scan the Internet incrementally, but the Internet cannot do it, with millions, or even hundreds of millions of cameras, as well as audio and other sensors in the real world.
I think we may have the most data and may have the greatest training computing power. Therefore we may contribute to general AI AGI.
Q12: We did not talk about the electric truck Semi. I want to know, from a perception point of view, what changes are you considering Semi? Compared with cars, the demand is obviously very different. If you don't think so, what's the reason?
Tesla electric truck Semi appeared at the AI day site, but was not specifically mentioned
Musk: I think, no matter what vehicle you drive, what are you needed? What is needed is that biological neural networks and eyes are basically cameras. Your main sensor is two cameras installed on the slow gimbal. The very slow gimbal is your head.
If a biological neural network plus two cameras on a slow gimbal can drive a Semi truck, then if you have 8 360-degree cameras looking around the vision running at higher frame rates and faster reaction speeds, I think it's obvious that you should be able to drive Semi or any vehicle better than humans.
Q13: Assuming Optimus Optimus will be used in different scenarios and evolve at different speeds for these scenarios. Is it possible to independently develop and deploy different hardware and software components for it and deploy them on Optimus Optimus? In this way, will Optimus Optimus develop faster in the future?
Musk: We don’t understand. Unfortunately, our neural networks fail to understand this problem.
Q14:I want to ask the autonomous driving system, when do you plan to promote the FSD beta version to countries outside the United States and Canada? What do you think is the biggest bottleneck or obstacle in the current technology stack of autonomous driving systems? How will you plan to solve the problem so that autonomous driving systems can surpass human driving in terms of performance, safety assurance, and confidence in human use? I remember you also mentioned that you planned to merge highways and cities into a unified technology stack in FSD V11 and make some architectural improvements. Can you explain it in detail?
Musk: From a technical perspective, the FSD beta version should be launched globally by the end of 2022, but for many countries, we need to obtain approval from regulatory authorities. To some extent, we are subject to regulatory approval restrictions in other countries. I think from a technical point of view, the beta version can be prepared to be launched worldwide by the end of 2022.
We expect significant improvements in October 2022, which will be particularly good at evaluating fast moving traffic vehicles, as well as other improvements.
John: In the past mass-produced version of the autonomous driving system and the FSD beta version, and these differences have become smaller and smaller as time goes by. I remember a few months ago, all FSD beta and production versions of autonomous driving systems used the same pure visual object detection technology stack.
There are still some differences between the two at present, the most important thing is how we predict lane. We upgrade the lane model, as I mentioned in my speech, so that it can handle more complex geometric shapes. In the mass-produced version of the autonomous driving system, we still use a simpler lane model.
We are expanding the current FSD Beta model to handle all highway scenarios.
Musk: I drive a Tesla car and use the FSD Beta version. In fact, it has a unified technology stack. It uses the FSD technology stack on both city streets and highways. For me, it performs very well.
We need to verify it in various weather conditions, such as heavy rain, snow, and dust, to ensure that it can perform better than the mass-produced version technology stack in various environments.
We are very close to this goal, and we will definitely be ready by the end of 2022, and it may also be November 2022.
Paul: Based on my personal driving experience, the highway FSD technology stack has far exceeded the mass-produced version technology stack. We hope that by the end of 2022, the parking technology stack will also be included in the FSD technology stack. Basically, before the end of 2022, if you sit in a car in the parking lot, the car will drive to a parking space at the end of the parking lot.
Musk: The basic indicator that needs to be optimized is how much mileage the vehicle can travel before necessary intervention. Significantly increasing the mileage of a vehicle that is fully autonomous before intervention is crucial to safety. This is the basic indicator of our weekly measurements and we have made huge improvements in this regard.
Q15: I am very curious, if you go back to your 20s, what do you hope you will know at that time, and what advice would you give to your younger self?
Musk: I'm trying to think of something useful to say...yes, joining Tesla is one of the things.
I want to try to get in touch with as many great people as possible. I read a lot of books, that's what I do.
I think it doesn't necessarily need to be too volume, there are some benefits to doing this. For me at the age of 20, it might be a good idea to enjoy the moment more and stop occasionally to smell the fragrance of flowers.
For example, when we were developing the Falcon 1 rocket, we developed the rocket on this beautiful island of Kwajalin Atoll (the largest island in the Marshall Islands in the Western Pacific). I didn't even drink a single drink on the beach during that whole time. I would say I should be able to have a drink on the beach and that wouldn't be a problem.
Falcon 1, which is ready for test launch in the temporary launch site of Kwajalin Atoll,
Q16: You use Optimus Optimus to make all robot industry practitioners excited. It feels a lot like the autonomous driving technology 10 years ago, but autonomous driving has proven to be much more difficult than it seemed 10 years ago. Is there anything we didn’t understand 10 years ago but now we can make the robot come sooner?
Musk: In my opinion, general artificial intelligence AGI is developing rapidly, and there is no shortage of important progress almost every week.
At present, AI seems to be able to win in almost all rules-based games, create impressive art, participate in very complex dialogues, write articles, etc. These are constantly improving, and there are still so many talented people studying AI, and the hardware is getting better and better.
is not only our work at Tesla, but AI is on a strong exponential development curve, and it is obvious that we will benefit from it.
Tesla happens to be very good at drives, motors, gearboxes, controllers, power electronics, batteries, and sensors.
As I said, the biggest difference between a robot on four wheels and a robot with hands and feet is to correctly solve the driver problem. This is a question about drivers and sensors, about how to control these drivers and sensors. You have to have all the elements you need to produce a competitive robot, and we are doing that.
Q17: Tesla and you are taking humanity to the next level. You said Optimus Optimus will be used in the next Tesla factory, and my question is, will the new Tesla factory be completely managed by Optimus Optimus? When can people order humanoid robots?
Musk: Yes, I think when starting, we will have Optimus Optimus perform very simple tasks in the factory, such as loading parts. As you can see in the video, transport parts from one place to another, or load parts into our more traditional body welding robot units.
At the beginning, we will try to study how to make it work, and then gradually expand its scope of use.I think Optimus Optimus usage scenarios will grow exponentially and will be very fast.
As for when people can order, I don't know, I don't think it's far away. I think you mean when people can receive robots. I don't know, I will say that it may be 3 years, no more than 5 years, within 3~5 years you may receive Optimus Optimus.
Q18: I think the best way to promote the progress of general artificial intelligence AGI is to let as many smart people as possible participate in the world. Compared with robot companies, considering Tesla's scale and resources and considering the current research status of humanoid robots, isn't it very reasonable for Tesla to open source some simulation hardware? I think Tesla can become the dominant platform provider and the Android or iOS system for the entire humanoid robot research field, rather than just allowing Tesla researchers or their own factory to develop Optimus Optimus, which can open up Optimus Optimus and allow the whole world to explore the research of humanoid robots.
Musk: I think we have to be careful that Optimus Optimus is likely to be used in a bad way, which can happen. You can issue instructions to Optimus, but these instructions are restricted by robot regulations that you cannot violate and cannot cause harm to others. I think Optimus Optimus may bring some security discussion.
Q19: What is the current and ideal controller bandwidth for Optimus Optimus? This event is a powerful advertising for the depth and breadth of the company. What is unique about Tesla that can do this?
David:For bandwidth issues, you must understand or figure out what you want to accomplish the task. If you accept a frequency conversion task, what do you want your limb to do? That's where your bandwidth comes from. This is not a specific number you can directly say, you need to understand your use case, and that is where bandwidth comes from.
Musk: Regarding the bandwidth issue, I think we may eventually increase bandwidth, and the effect is equivalent to improving robot flexibility and reaction time. You can save the robot state, not at 1 Hz speed, but you don't need to increase to 100 Hz level. I don't know what it is, maybe 10 Hz or 25 Hz.
Over time, I think the bandwidth will increase significantly, or it is equivalent to improving flexibility and reducing latency. We hope to minimize latency and maximize flexibility over time.
We are currently a considerable company, and we have many professional and technical developments in many fields to develop to develop electric vehicles and autonomous driving technologies. Basically, Tesla is a combination of several startups, and so far they are almost all successful, so we must have done something right.
I believe that one of the core responsibilities of management companies is to provide an environment for great engineers to grow fully. I think in many companies, maybe most companies, if someone is a really talented engineer, their talent is suppressed in many companies.
In some companies, the way talent is suppressed may not look that bad, because it looks comfortable to work, your salary is high, but the output you need to achieve is so low, which is like a "sweet trap".
In Silicon Valley, there are some "sweet traps". They don't look like bad choices for engineers, but you have to say, a good engineer joins, what output they achieve? The output of these engineering talents seems to be low, even if they seem to be having a happy life. That's why I say there are some "sweet trap" companies in Silicon Valley.
Tesla is not a "sweet trap". We have high requirements. You need to complete a lot of work, and those jobs are cool and not simple.But if you are a super talented engineer, your talents will be more fully utilized than elsewhere. The same is true for
SpaceX.
Q20: The first question, I have been following your progress in the past few years. Today you have made some changes in lane inspection. You said that you used instant semantic segmentation before, but now in order to create lane information, you have established a transfer model. What are the common challenges you still face? For example, as curious engineers, we as researchers can start studying these problems.
The second question, I am very interested in the data engine. You show a case where the car stops driving. How did you find very similar examples from the data you have? It would be great if you could introduce the data engine.
Phil: I will answer the first question first. Take "occupying the network" as an example. You see content in your speech, which did not exist 1 year ago. We only spent 1 year to release more than 12 "occupying network" models.
is a basic model to express the entire physical world under all places and all weather conditions, which is very challenging.
Just over a year ago, we only drove in the two-dimensional world. We used the same static edge to represent the wall and curb, which is obviously not ideal. There is a big difference between the wall and the curb, and when you drive, you make different choices.
After we realized that we had to rethink the whole problem and consider how to solve it. This is an example of the challenges we have to overcome in the past 1 year.
Kate: Back to your second question, how do we actually get tricky examples of parking vehicles, there are several ways to solve it, give two examples.
first, we can trigger divergence signals. The bit that indicates parking will flash between parking and driving states, which will trigger data return.
Second, we can better utilize the logic of the shadow pattern. If the customer ignores a car but we think it should stop because of this, we will also pass this data back.
These are different trigger logics, allowing us to return and get this data.
Q21: There are many companies that are paying attention to the general artificial intelligence AGI problem, which is a very difficult problem. One of the reasons is that this problem itself is difficult to define. Different companies have different definitions, and their focus angles are also different. How does Tesla define general artificial intelligence AGI problems? What are your specific concerns?
Musk: We are not actually paying special attention to the general artificial intelligence AGI problem. I'm just saying that general AI AGI seems likely to be an emerging attribute of what we're doing.
We are creating all these self-driving cars and humanoid robots that are in a huge data stream, data flowing in, processed, which is the largest real-world dataset to date.
These data, you cannot just search the Internet, you must go into the outside world, interact with people, and interact with the road. The earth is a huge place, with chaotic and complicated reality.
I think this seems to be an emerging property. If you have tens of millions or hundreds of millions of self-driving cars, or even a considerable number of humanoid robots, maybe there are more humanoid robots, then this is the largest number of data sets.
These videos are processed and it is very likely that cars will definitely be much better than human drivers, and humanoid robots will likely become increasingly difficult to distinguish from humans.
As I said, you have the emerging attributes of general artificial intelligence AGI.It can be said that all human beings are also super intelligent as a whole, especially when we increase the data interaction rate between humans.
Similar things seem to have happened a long time ago, and the Internet is like the nervous system acquired by humans. Suddenly, by connecting to the Internet, any part of humanity can master all knowledge, almost all knowledge, at least a large part of it.
Before, we exchanged information through individual penetration. In order to transfer data, you have to write a letter, someone has to be responsible for handing the letter to another person, and there is a lot of things happening in the middle. Think about it, it's really slow.
Even if you are in the National Library, you still cannot obtain all the information in the world, and you will definitely not be able to search for this information. And it is obvious that not many people will be in the National Library.
I mean, one of the great equalization elements, the Internet is the largest equalizer in history in terms of obtaining information or knowledge. I think any student studying history would agree with this.
Recently someone asked me, which historical period I like most? My answer is now. Now is the most interesting time in history, and I have read a lot of history books to keep us doing our best to keep this.
Back to the previous question, my answer is, regarding Tesla's autonomous driving system, over time, neural networks gradually digest more and more parts of the software.
Of course, in extreme cases, you can simply get the video you see in the car and compare it with the steering input of the steering wheel and pedal. These are very simple inputs. In principle, you can train directly between the two because that is what humans are doing with biological neural networks.
training video shows the steering wheel and pedal moving, and there is no other software in the middle. We have not reached this level yet, but we are gradually moving in this direction.
Q22: So many people use FSD, in terms of performance statistics, how do you evaluate the company's risk tolerance? Do you think there is a need for greater transparency, or from third-party regulation, to determine what is good enough and to define performance thresholds based on such a large amount of miles?
Musk: Tesla's primary design requirement is safety, which is the case for all aspects.
In terms of automotive machinery safety, we are among all the cars tested by the government, and in terms of passive machinery safety, the probability of injury is the lowest. We also received the highest rating in terms of active safety.
In the future, it will reach this level: the security is surprisingly high, even much better than humans.
Regarding the autonomous driving system, we have indeed published statistical data in a broad sense. According to the mileage statistics, Tesla cars do not turn on the autonomous driving system, and compare the situation of turning on the FSD beta version. We have seen that along the way, the stable improvements of the autonomous driving system have been made.
Sometimes there is such a contradiction: Should we wait until the car is three times safer than humans before deploying this technology?
But I think this is actually a wrong idea in morality. As long as you believe that autonomous driving can reduce harm and death, you have a moral obligation to deploy it, even if you will be prosecuted and blamed by many people.
Because those who have been saved by you do not know that their lives have been saved. Those who occasionally die or get injured, they certainly think that, in any case, it is the fault of the autonomous driving system.
This is why you have to look at the total mileage statistics, how many accidents occur, how many accidents are serious, and how many deaths are caused.
We have more than 3 million cars on the road, and the total mileage is very high every day. It won't be perfect, but importantly, it's obviously safer than not deploying.
Q23: I don’t work in hardware. Maybe the hardware team and you all can give me some inspiration. Why does Optimus Optimus design need to have symmetry? Because we humans have left and right hands, right? We use certain muscle frequency, higher than others, and wear and tear over time. So you may see that some joint failures, or some actuator failures occur more frequently.
As time goes by, I understand that it is still very early. As humans, we have a lot of fantasy and fiction about the superhuman abilities. All of us don’t want to go there directly. We want to stretch our arms and want all these fantasy designs. With everything else in mind, maybe you can take advantage of all these aspects to create more interesting robots in terms of battery and computing strength. I hope you can explore these directions.
Musk: I think it would be cool to turn Detective Gajet (the protagonist of the American cartoon "Detective Gajet") into real, that would be great.
At present we just want to produce a humanoid robot with basic work. Our goal is to produce useful humanoid robots at the fastest speed. I think this will put us on reality and make sure what we do is useful. One of the most difficult things about
is to be useful and then achieve the total value of high efficiency under the curve. For example, how much help you provide to each person on average, multiply by how many people you help in total, and get the total utility value.
strives to release useful products that people like to a large number of people. It is extremely difficult and the degree of difficulty is unimaginable. That's why I say that a company with released products is a huge difference from a company without released products, which is simply a world of difference.
Even if you release a product, can the output value exceed the input cost? This is also extremely difficult, especially for hardware.
I think it would be cool to do something creative over time, such as having 8 arms, publishing different versions. Maybe some hardware companies can add some features to Optimus Optimus, say, maybe a power interface, or something like that. Or you can add attachments to Optimus Optimus, just like adding attachments to your phone.
There may be a lot of cool things to do over time. There may be an ecosystem of many small companies, or even large companies, to produce accessories for Optimus Optimus.
Conclusion
Musk: I want to thank the team for their hard work, you all are amazing. Thank you for coming, and thank you everyone online for your attention.
I think this will be a great video and you can also fast forward to the part you think is the most interesting. But we strive to provide you with a lot of details so that you can watch videos in your spare time. You can focus on the parts you find interesting and skip the others.
Thank you everyone, we will work hard to hold such events every year, we may even do podcasts once a month, but it is great to get everyone involved in this process and show you the cool things that are happening.
"Long-termism" column is updated every Saturday and long holiday, divided into the following series:
Macro theory: politicians, business leaders, etc. in major countries around the world, etc.
Social theory: Bill Gates and other
Social theory: Bill Gates and other
Growth theory: Rockefeller, Carnegie, etc.
Science theory: Nobel Prize winners, Tencent Science WE Conference, etc.
Science theory: Musk, Bezos, Larry Page/Sergey Brin, Zuckerberg, Huang Renxun, Vitalik Buterin, Brian Armstrong, Jack Dorsey, Son Masayoshi, Huawei, Ma Huateng, Zhang Xiaolong, Zhang Yiming, Wang Xing and others
investment statement: Buffett, Munger, Baillie Gifford, Howard Max, Peter Till, Mark Anderson, Katherine Wood, etc.
Management theory: Ren Zhengfei, Kazuo Inamori, etc.
Energy theory: Zeng Yuqun, etc.
l3 Auto says: Li Xiang, He Xiaopeng, Wang Chuanfu, Wei Jianjun, Li Shufu, etc.
Intelligent says: DeepMind, OpenAI, etc.
Meta Universe says: Meta/Facebook, Apple, Microsoft, Nvidia, Disney, Tencent, ByteDance, Epic Games, Roblox, Bilibili/B station, etc.
Interstellar: China National Space Administration, NASA, International Aerospace Conferences over the years, SpaceX, Starlink, Blue Origin, Virgin Galaxy, etc.
Consumption: Amazon, Walmart, Alibaba, JD.com, Pinduoduo, Meituan, Oriental Selection, etc.
Each series focuses on top global experts and industry leaders in various fields, collects and organizes first-hand information such as shareholder letters, public speeches/communications, media interviews, etc., learn classics together, absorbs ideological nourishment, be friends of time, and be long-termists.
Related research reports:
[Presale of the No. 1 Player Season 2]: Technological innovation brings paradigm transformation, expanding the endless new frontier
Jiuyu Capital Zhao Yujie: 15,000 words Number One Player's annual thinking set, technological innovation, endless frontier
Jiuyu Capital Zhao Yujie: 15,000 words Number One Player's annual thinking set, technological innovation, endless frontier
Jiuyu Capital Zhao Yujie: 15,000-word smart electric vehicle annual thinking set, software definition, reshape everything
[Big] Frontier Weekly Report: Embrace technology and see the future, 70th issue collection packaged and sent to
[Big] Liuhe Annual Report Full Library Members are officially launched, and more than 5 years of research results are systematically delivered
[Smart Electric Vehicle Special Presale] A century-old automobile industry accelerates its transformation, smart The curtain of the electric vehicle era begins
[Presale of the first season of the first player]: Technology giants explore the future, the rise of the number one player echelon
[First paid report + annual membership] Live e-commerce 140,000 words in-depth report: under the trillion-level GMV trend, giant melee and alliance with the alliance
[Most] The collection of science and technology sports series reports is online, "Technology + Sports" is deeply integrated , All-round transformation of sports
[Big Up] 365 star companies, nearly 600 reports, Liuhejun's more than 4 years of research results are panoramic presentation of
Jiuyu Capital Zhao Yujie: CES experiences, open a brain, super technology giants will take over everything
[Ten Thousand Words long article] Jiuyu Capital Zhao Yujie: 5G opens a new cycle and enters the era of sailing in the online world | GBAT 2019 Greater Bay Area 5G Summit
Jiuyu Capital Zhao Yujie: Grasp the huge trend dividend of e-cigarettes, seize the change and unchanging in industrial transformation
[IPO Observation] Season 1: SMIC, Cambrian, Smoore, Pop Mart, Anker Innovation and other 11 in-depth research reports collection
[IPO Observation] Season 2: In-depth research reports of 12 companies including Ideal, Xiaopeng, Shell, Ant, Snowflake, Palantir Collection of
[IPO Observation] Season 3: Coinbase, Roblox, Kuaishou, Wuxin Technology and other 12 companies in-depth research reports collection
[Big Up] Annual Observation 2019 Series Collection: Over 3 months, more than 200,000 words, nearly 500 pages, review the past, look forward to the future, insight into change and unchanging
[Collective Edition] Liuhe Book: Panoramic scan of 300 celebrity companies, lasting 3 years, more than 5,000 pages of 2.1 million words , Recommended
Jiuyu Capital Zhao Yujie: Thinking about the fragmented smart electric vehicle industry
Jiuyu Capital Zhao Yujie: Nine-grid analysis method, Taoism and techniques in the field of Chinese, mathematics and foreign education and training
[2022 Returning to Hometown Experience] 20 90 and 00 post-00 20,000 words, 4 countries and 13 places, showing the differences in epidemic prevention, Spring Festival atmosphere, and development status at home and abroad
[2021 Return to Hometown Experiences] 22 post-90s 20,000 words, celebrating the New Year at home and celebrating the Chinese New Year collision, showing the real, three-dimensional and transformative Chinese
[2020 Return to Hometown Experiences] 20 post-90s 20,000 words, special Spring Festival, collective memory of the times
[mainly] 22 post-90s 20,000 words return to Hometown Experiences, telling the story of China's counties, towns, and rural areas in their eyes
Liuhejun's 3rd birthday, TOP 60 classic research reports recommended
afternoon tea, the Three Kingdoms Kill in the Internet world
new commercial infrastructure continues to be enriched and improved, empowering the rise of new brands, new models, and new products , create a new diversified lifestyle
[Big Up] China's new economy leader, 20 listing reports on Hong Kong and the United States
Knowledge service + payment + audio, opening up new industrial opportunities for content production, the knowledge economy has a market size of hundreds of billions in 10 years, which can be expected to be
from APP Store best-selling list has changed for 4 years, and the rise of content payment is
New Third Board, we have a Jiuyuhui family
New Third Board has exceeded 10,000 Thoughts: The daily trading volume of the New Third Board in 10 years is 100 times?
Jiuyu Capital Zhao Yujie: Technology changes consumption, making life better | 2017 GNEC New Economy New Consumption Summit
Jiuyu Capital Zhao Yujie: The law of time in the era of entrepreneurship, develop user parallel time | 2016 GNEC New Economy New Intelligence Summit
Jiuyu Capital Zhao Yujie: The Internet leads the new economy, and content entrepreneurship connects to a new ecosystem | 2016 GNEC New Economy and New Marketing Summit
Please be sure to read the disclaimer and risk warning
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-