On October 1, 2022, on Tesla's new AI Day, Tesla once again showed off humanoid robots. Compared with the cool dance performance last year, this year's one seems a bit staggering.

[Pacific Auto Network New Car Channel] On the Tesla AI Day last year, a humanoid robot danced wildly to spark heated discussion. Of course, that flexible robot is a conceptual display of live actors, and everyone thinks it is just a muscle-showing performance.

On October 1, 2022, on Tesla's new AI Day, Tesla once again showed off humanoid robots. Compared with the cool dance performance last year, this year's one seems a bit staggering. Unlike last year, the Tesla Bot humanoid robot "Optimus" prototype that took to the stage today does not have a shell, and joints, bones, cables and other equipment are clearly visible, and each action details are visually displayed. In just 6 months, Tesla Bot quickly moved from one concept to reality.

, and Musk said that Tesla's humanoid robots will be mass-produced in as early as three years. Of course, for Tesla, what we are currently paying more attention to is still his automotive business, and the latest progress on FSD (full autonomous driving capability) and Dojo supercomputers have also been announced on 2022 AI Day. The Optimus prototype is also closely related to Tesla FSD and Autopilot-related neural network technologies.

Science fiction enters reality. Humanoid robots replace artificial?

Turning around, stopping, and waving greetings, the Optimus prototype completed a standard humanoid robot debut ceremony. In the video on the press conference, Tesla showed a series of usage scenarios such as Optimus lifting a kettle to water flowers in the office, carrying materials with both hands to the target position, accurately locate people around him and actively avoiding him. Even Optimus has started to put in simple work at Tesla's Fremont factory and planned corresponding courses of action for it.

official introduction, Optimus can currently complete movements such as walking, going up stairs, squatting, and picking objects. The flexible hands benefit from reasonable finger joint settings, and have excellent flexibility and strength performance. The actuator can not only withstand heavy objects like a piano weighing about half a ton, but also complete high-precision movements such as light and thin objects grabbing, operating mechanical devices, and complex gestures.

Optimus prototype is equipped with a 2.3 kWh, 52V battery pack at the torso, which is highly integrated with charging management, sensors and cooling systems. "This means that from sensor integration to charging management, it is all brought together into this system, which is also a reference to our experience in automotive design." Tesla engineers said. In terms of structure, its body has a total of 28 degrees of freedom, adopts a more flexible spring load design and 6 types of actuators. The joints are designed with bionic joints, which simulates the shape of human joints and tendons, and the hands are ergonomic design, with 11 degrees of freedom.

. In terms of "sensory", Tesla robots use the boundaries of objects recognized in full vision and use the Tesla model's FSD supercomputing platform for training. Walking may be a simple daily movement for humans, but robots have great challenges to achieve stable walking. First of all, Tesla needs to give robots a physical self-awareness, ensure exercise energy efficiency, maintain balance, and quickly respond to risks in the environment, while also taking into account limb length, weight and other aspects.

At the same time, Tesla also creates robot safety based on its automotive safety simulation and analysis capabilities. In traffic accident simulation, Tesla improves safety performance through software optimization and improvement of vehicle collapse and battery protection. In robot design, Tesla also ensures Optimus' ability to protect itself and people around it in the same way. For example, in external situations such as falls and collisions, the robot will make decisions consistent with humans - giving priority to ensuring the safety of the "brain" to the greatest extent, followed by the safety of the trunk battery pack. The importance of these two parts is self-evident.

At present, Optimus can not only walk smoothly, but also maintains balance with a small amount of external interference. In the live video, Tesla also showed Optimus being poked by a wooden stick by a staff member, but did not fall or shake for a long time. He just leaned forward and stabilized his posture, and responded to the challenge excellently.

"When talking about robots, you will think of economic development.The fundamental element of the economy is labor. If we can use robots to achieve lower labor costs, it will eventually lead to faster economic development. "Musk said that in the future, with the participation of robots, people can freely choose their work, among which simple and repetitive work can be handed over to robots. This type of work will become a choice for humans, rather than a need.

Musk believes that after solving the problems related to the core of robots, it can further "add icing on the cake". In the future, Optimus can not only complete tasks very effectively" after evolving, but also become our partner. People can take it out for food, drink, play, wear various clothes for it, and choose a variety of ways to get along with it to form a unique personality. "I believe that under everyone's creativity, the colorful Optimus will be created. "We should be able to produce millions of units, and it is expected to be much cheaper than cars, and the robot is estimated to end up at under $20,000. "He also predicted that Optimus will be mass-produced and launched in 3-5 years.

What's more interesting is that Musk immediately said on social media that Tesla Bot will definitely launch the "Catwoman" version of the robot. Are you looking forward to it?

uses FSD training experience to efficiently build the robot "brain"

The brain of the machine is artificial intelligence . Optimus is equipped with the same FSD computer as Tesla vehicles and Autopilot-related neural network technology. It uses similar full-visual boundaries of object recognition. Therefore, the research results in autonomous driving technology have also helped Tesla lay a solid foundation in the field of humanoid robots.

"In fact, Tesla vehicles have installed FSD computers. We hope to combine autonomous driving experience with humanoid platforms," ​​said Tesla engineer. "But the actual needs and forms of robots are different from those of cars, so we still need to adjust appropriately. "

Tesla hopes that the robot's "brain" should do everything that the human brain can achieve as much as possible, such as processing visual data, making action decisions based on multiple sensor inputs, supporting communication and communication, etc. Optimus uses the same chip as the Tesla vehicle, which supports Wi-Fi, LTE links and audio communication, and the security guarantee of the system's software and hardware is also improved again.

At the same time, Optimus also "learned" humans through motion capture . To move As an example, Tesla staff enter actions through wearable devices, and the robot learns through neural network , from completing the same action at the same location to evolving to deduce the scheme in other scenarios, thus learning to carry different items in different environments.

FSD and Dojo supercomputers announced the latest progress, and the training efficiency is greatly improved by 1,000 times. In addition, Tesla demonstrated the progress of FSD (full autonomous driving capability). Based on the previous basis, the recognition ability and execution decisions were optimized. The Beta version can already drive automatically from the parking lot to another parking lot, pass the intersection, and wait for the traffic light . Now the FSD Beta can predict the surrounding environment and act, predict the next action of other vehicles and pedestrians on the road, and give corresponding instructions to the vehicles.

In the unprotected left turn scenario, the latest FSD Beta will slow down first, choose a relatively safe route, be more personified, and even pass through the gap between vehicles and pedestrians. Previously, the FSD Beta would park at the intersection at a relatively fast speed, and wait for the vehicle to pass before passing. This upgraded trajectory judgment requires a lot of simulation calculation and simulation. Tesla has applied neural network learning to it and has done a lot of work and analysis to ensure that the vehicle passes safer and more comfortable. FSD Beta's deep neural network reaches 1 million parameters, 150,000+ neural network layers, and 375,000 connections. It can simulate scenes that have not appeared in reality within 5 minutes.

Tesla is currently deploying the FSD Beta version widely in North America and will launch large-scale testing of the fully autonomous driving capability (FSD) software this year. The test users have expanded from 2,000 in 2021 to 160,000, and 75,000 neural network models have been trained in just one year.The cumulative mileage of the beta version of the software has exceeded 40 million miles, and it is expected to exceed 100 million miles by the end of the year.

In the year, the number of participants in the Tesla FSD Beta test version has increased by 80 times. "In 2021, 2,000 cars participated in the testing of the FSD Beta version software, and we have expanded this data to 160,000." Tesla engineers introduced that in 2021, Tesla trained a total of 75,000 neural network models, which is equivalent to training one every 8 minutes and pushing 35 version updates. Now the FSD Beta version of the software can realize fully autonomous driving to a certain extent, such as navigating from one parking lot to another, braking and waiting for traffic lights, coordinating at intersections and other targets, etc.

Tesla's efficient data training is due to the Dojo supercomputer first launched on AI Day last year. The single training module of the Dojo supercomputer is made of a whole wafer, and each training module contains 25 Tesla's independently developed D1 neural network training chips. Since the entire wafer is not cut, each D1 chip is seamlessly connected together, and the delay between adjacent chips is extremely low. The training module achieves bandwidth retention to the greatest extent, and is combined with Tesla's own high-bandwidth and low-latency connectors.

Dojo single chip computing power is 362 TFLOPS, and 25 chips are in a group to form a training module. The computing power can reach 9 PFLOPS and the interface bandwidth is 36 TB/s. The cabinet model launched by Tesla this time consists of 120 training modules, built-in 3,000 D1 chips, more than 1 million training nodes, and the computing power will reach 1.1 EFLOPS, which is ten thousand times that of Tesla's existing supercomputers. And Dojo supports unlimited connections, and theoretically, there is no limit on performance. According to public data, Dojo's computing power has reached fifth in the world, second only to Japan's "Fugao" in the first place, the US Summit in the second place, the US Sier in the third place, and the China's "Shenwei Taihu Lake" in the fourth place.

Tesla does not set limits on Dojo's design, it can provide a very large and efficient system. After the Dojo supercomputer was put into use, Tesla has achieved a 30% training speed improvement in code and design. For example, through automatic labeling technology, Tesla has increased the labeling speed of training scenarios by 1,000 times. Under manual operation, these may take one or two weeks or even several months.

Under efficient neural network training, Tesla FSD software has gradually matured at the technical level. Musk said: "At present, FSD software can be applied to road conditions in various regions around the world. If local regulatory policies allow it, we can launch the FSD Beta version of the software globally by the end of this year."

Tesla said it will arrange the first batch of 10 Dojo supercomputer cabinets in the first quarter of 2023, that is, ExaPOD with computing power exceeding 1.1EFLOPS, and the automatic labeling capacity will be increased by 2.5 times; in the future, Tesla will deploy 7 such clusters in Palo Alto to provide unimaginable huge computing power. At the same time, Musk also said that Dojo supercomputers will be provided to other users on the cloud in the future like the Amazon web service.

Conclusion

2021 On AI Day, a humanoid robot played by a live dancer and a just completed Dojo supercomputer sparked heated discussion. Just a year later, Tesla Bot robots have truly stepped onto the stage. The first batch of 10 Dojo supercomputer cabinets with computing power exceeding 1.1EFLOPS are about to be deployed, which makes people sigh at Tesla's amazing strength. There are still some variables when Tesla Bot robots can be truly mass-produced, but Musk’s bold creativity to let robots liberate human productivity is still exciting. (Text: Pacific Auto Network Guo Rui)