"The future price of Tesla robot is US$20,000 (equivalent to RMB 140,000), far lower than the price of Tesla car ".
"With Tesla robots, poverty can be eliminated, and humans can freely choose their own profession."
This is this year's Tesla AI DAY. While taking out the humanoid robot, Musk described to the outside world a beautiful future world where humans and robots coexist harmoniously.
FSD fully autonomous driving and the latest progress of Tesla's self-developed supercomputer Dojo are the other two major themes of AI DAY this year.
Of course, this AI DAY still remains the same as the "recruitment conference" in the past, and Musk's recruitment order runs through the end from the beginning.
In short, all Tesla’s practical information and unfilled pitfalls in the AI field this year have been fully displayed on AI DAY.
Cyber Auto will take you to read all the articles.
Humanoid robot walked out of PPTh
The last AI DAY Tesla Bot, which debuted as the finale, was the first to debut this year, and it is the following:
, the above is the verification machine, and the following is the main owner. Bot Generation:
The prototype unveiled this time is already somewhat similar to the robots on PPT, but the torso design looks more mechanical.
At the same time, it can also be seen that the Tesla Bot prototype exhibited this time has the ability to move independently, and the flexibility of the limbs is also good. It is still guaranteed to do a simple dance move:
Of course, it would be too bad if you just have a simple activity. This time, more and more detailed technical parameters of Tesla Bot are announced at the same time:
First of all, the weight of the Tesla Bot generation prototype weighs 73 kilograms, an increase of 17 kilograms from the original official data.
Body flexibility data is also good. Tesla Bot's whole body freedom can exceed 200, and the palm freedom can reach 27.
Musk said that the robots on display can now realize the ability to move independently and operate tools, and do some simple and repetitive work, while the palm can carry 20 pounds (about 9 kilograms) of weight.
In terms of body layout, most of the Tesla Bot actuators are concentrated in the limbs. The power battery system is laid out on the torso, equipped with a 2.3kWh battery pack, with a nominal voltage output of 52V.
computing hardware is also located in the torso part. Tesla Bot is equipped with an FSD chip of the same model as Tesla, supporting Wi-Fi and 4G (LTE) network transmission.
In addition, Tesla Bot also embeds Tesla's technology in autonomous driving, including the software algorithm and pure vision solution of the FSD autonomous driving system .
In addition to the above parameters, more R&D processes and technical parameters behind Tesla Bot have also been disclosed:
In terms of safety performance, Tesla Bot applies some of the technology of Tesla's car impact experiment to ensure that the robot will not have major damage when it falls or is hit (the battery pack is on the chest, and it is not good if a robot spontaneously ignites).
At the same time, considering Tesla Bot's own weight and purpose, the flexibility and pressure-bearing ability at the joints also need to be taken into account. In this regard, Tesla Bot's joint design references human skeletal tissue.
In order to ensure the pressure and flexibility of the 28 joint actuators in the whole body in different states, while taking into account efficiency and energy consumption, Tesla simulated through point cloud and designed 6 actuators with different torque outputs to undertake mobile tasks.In terms of software details, Tesla mainly emphasizes the following parts:
First of all, as a humanoid robot, it needs to face various risks in the physical world. Therefore, Tesla officially emphasizes that Tesla Bot needs to have a physical level of autonomy, such as understanding of its trunk body and the real world.
Secondly, it is dynamic stability. Tesla achieves walking balance and path planning through a lot of simulation training. After adding the navigation system, Tesla Bot can move independently and even find a charging station to replenish its energy.
Finally, Tesla Bot's movement details are anthropomorphized. Tesla visualizes the human movements, analyzes the position of the trunk, and maps it to Tesla Bot.
After talking about the technical details of Tesla Bot, let’s try the big cakes drawn by Musk (I respect Musk as the first person to paint the cakes).
Musk said that Tesla Bot will be more expensive in the early stage, but as the output increases, the price will also drop sharply, and it is expected to reach a million-level output in the future. In the end, the Tesla Bot will be priced at a lower price than the Tesla electric car, which is expected to reach US$20,000 per unit (equivalent to about RMB 140,000). How about
? Is a robot tempted by $20,000? If you can’t buy it, you will suffer losses, you will be fooled~
At the same time, Musk also said that with robots, the economic situation will be better and poverty will be eliminated. Because robots can replace cheap repetitive labor, future people can choose their favorite jobs independently. Is the screen of
a bit familiar?
But! Don't worry, the pictures described by Musk and most technical details, including navigation, flexibility, etc., still require a lot of refinement.
As for when it will be realized, Tesla officially stated that it may take several weeks, one year, or several years...
Yes, Musk dug a hole again, Tesla Bot, after all, can't get rid of the same fate as FSD.
But you can still look forward to it here. After all, most of Musk's promises appear in the form of "although it's late". (You can add a dog head here)
FSD is still difficult to get to this part of Beta
FSD, there are not many surprises. The mass-produced version of FSD is still far away, and the expected Beta v11 version has not arrived either.
However, from the technical introduction, Tesla has started to specifically target the long-tail issue (Corner Case). The autonomous driving team interpreted the technology stack in detail, as if saying, "You see, autonomous driving is so difficult to achieve, and I can't blame me for always skipping the tickets."
Through this AI Day sharing, we can see that Tesla has built an autonomous driving technology stack that supports FSD's rapid growth.
In 2021, 2,000 Tesla owners participated in the FSD Beta test, and by 2022, there have reached 160,000 owners participating in the test.
Behind every autonomous driving decision is a balance based on many factors. Whether the planning choice is radical or conservative involves the situation of the vehicle itself, and also involves other factors of traffic participation. This includes a lot of "relationship" processing, and there are many different states for different object indicators, including speed, acceleration, static, etc., which requires a large amount of edge computing support.
And as the traffic relationship increases, the calculation volume will become larger and larger. In a set of interactive relationships, all interaction indicators should be taken into account, and the most reliable solution should be calculated to form an decision tree .
At present, FSD has been able to shorten the running time of each operation to 100 microseconds, but Tesla believes that this is far from enough. In the future, analysis of riding comfort and external intervention factors will be added.
Tesla insists on a pure visual route, but the visual solution is not perfect. Although Tesla has 8 cameras, there will always be a certain blind spot in the actual traffic scene.
Therefore, in the neural network system of FSD, a space occupancy model is introduced. Based on geometric semantics, it analyzes how volume occupies by camera calibration, reducing latency, and renders it into vector space to represent the real world in a complete 3D way.
In addition, complex road structures are also an obstacle to the learning of autonomous driving. Human drivers need lane lines to guide their passage in the real world, and the same is true for autonomous driving.
The real world road structure is complex and has many connections, which greatly increases the difficulty of data processing. Tesla transforms these problems into semantics that can be recognized by computers through the lane detection network.
visual perception solution often generates a very large amount of data. The data processing compiler developed by Tesla is training video models through more efficient computing power utilization, introduction of accelerators, introduction of CPU, and reducing broadband loss, which has achieved a 30% training speed improvement.
In data annotation, Tesla says that data quality and labeling quality are equally important. It is currently adopting a combination of manual labeling and automatic labeling to achieve a more detailed labeling solution.
In the construction of simulation scenarios, Tesla has also made significant improvements in simulation scenarios. The simulation scenario generation is 1,000 times higher than before. The edge geometry modeling is more refined, and it can quickly copy the real environment and fully consider different driving scenarios.
Why Tesla does supercomputer
"Some people often say that as an autonomous driving company, why should Tesla develop supercomputer?"
This AI day, Tesla gave the answer: In essence, Tesla is a hard-core technology company.
"Ask this question because we don't know enough about Tesla and don't know what we are going to do. At Tesla, we do a lot of science and engineering-related things, so we have a lot of basic data work, including reasoning, neural networks, etc., and of course supercomputing."
After all, computing power can be said to be the basic food for training.
When designing the Dojo supercomputer in the beginning, Tesla hoped to achieve substantial improvements, such as reducing the delay in autonomous driving training. To this end, it has carried out a series of research and development, including high-efficiency chip D1, etc.
D1 chip was unveiled last year. It is a neural network training chip independently developed by Tesla. It is equipped with 50 billion transistors on a chip area of 645mm². The thermal design power consumption (TDP) is 400W, and the computing power peak under FP32 accuracy reaches 22.6 TFLOPS.
performance parameters are better than the Nvidia A100 Tensor Core GPU currently used by Tesla supercomputers. The latter has a chip area of 826mm², a transistor number of 54.2 billion, a TDP400W, and a peak computing power of FP32 is 19.5TFLOPS.
and the single training module of the Dojo supercomputer consists of 25 D1 chips. It is reported that Tesla will launch Dojo cabinets in the first quarter of 2023. At that time, the existing supercomputer built on the Nvidia A100 chip may be replaced.
In the future, data from more than 1 million Teslas around the world will be gathered in Dojo. Through its training of deep neural networks, it will help Tesla's Autopilot continue to evolve and ultimately realize fully autonomous driving (FSD) based on pure vision.
Tesla said that the new Dojo supercomputer has the advantages of artificial intelligence training ultra-high computing power, and it also has the advantages of expanding bandwidth, reducing latency, and saving costs.
Dojo team claims that the machine learning training computing power of a training module is enough to reach 6 "GPU computing boxes", and the cost is less than the "one box".
In order to achieve these performances, Tesla tried different packaging technologies and failed. In the end, Tesla gave up the D Ram structure and used S Ram, which is embedded in the chip. Although the capacity is reduced, the utilization rate is significantly improved.
In addition to architecture design, considering virtual memory, accelerators, compilers and other aspects, Tesla faces various choices in the entire system design, and they also follow their own pursuit, namely, "no limits on Dojo supercomputers."
For example, in the training method, most selected data co-line mode is not adopted; at the data center level, vertical integration structure is adopted to vertically integrate into the data center.
also encountered many challenges in this process.
Tesla hopes to improve performance by increasing density, which poses a challenge to power delivery. "We need to provide power and power for the computing chip, which will face limitations. At the same time, since the overall design is a highly concentrated, it is necessary to implement a multi-layer vertical power solution."
Based on the above two points, Tesla built a rapid iteration, and finally, through design and stacking, the reduction of CTE (thermal expansion coefficient) was as high as 50%.
Another challenge facing Tesla is: how to promote the boundaries of integration.
Currently, Tesla's power modules are x and y planes for high bandwidth communication, and everything else is stacked vertically. This not only involves the system controller, but also considers the loss of oscillator clock output. How to prevent it from being affected by power circuit and achieve an ideal integration level?
The method Tesla uses is a multi-pronged approach. On the one hand, it is to minimize vibration, such as by using soft cap ends, that is, the port uses softer materials to reduce vibration; on the other hand, update the switching frequency to further move it away from the sensitive frequency band.
At last year's AI day, Tesla only showed a few components of the supercomputing system, and this year it hopes to achieve more progress at the system level. Among them, the system pallet is a very critical part of realizing the vision of a single accelerator, and can achieve seamless connections as a whole.
In addition, in terms of hardware, Tesla also uses high-speed Ethernet, local hardware support and other methods to accelerate the achievement of supercomputing performance; in terms of software, Tesla said that the code runs on the compiler and hardware, and it is necessary to ensure that the data can be used together, so the path gradient needs to be considered in the reverse direction.
How to judge whether Dojo is successful and whether it has an advantage compared to the current situation? Tesla said it depends on whether colleagues are willing to use it. In fact, Tesla has also given some quantitative standards, such as the system's workload in a month, and the Dojo supercomputing can be completed in less than a week.
Of course, ultra-high computing power means huge energy consumption. During the Q&A session, Musk also said that Dojo is a giant computer, which consumes a lot of energy and requires a lot of cooling devices, so it may be provided to the market as an Amazon network such as aws and cloud service .
Musk believes that it is more meaningful to provide Dojo services similar to Amazon Cloud AWS. He describes this as an online service that helps you train models faster with less money." Tesla robot only sells more than 100,000 yuan, and FSD pushes 160,000 yuan to test
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-