When Optimus ("Optimus") humanoid robot appeared at the press conference, this means that another Musk flag was realized.
"The robot we are showing now is the robot we are about to put into use." Musk said that Tesla is confident of mass production at a lower cost, with a price expected to be less than US$20,000 (equivalent to RMB 140,000), which is lower than Model 3, and the final output will reach millions. In addition to bringing Optimus Prime humanoid robots, Tesla also announced the progress of FSD and Dojo supercomputers on AI Day.
It can be said that this is a press conference that brings together a number of hard-core technologies.
can carry and resist "Optimus Prime". It is expected to be priced at about US$20,000
Optimus Prime robot is about 170cm tall and weighs 73 kilograms. It is equipped with a 2.3kWh battery pack, runs on a Tesla SoC, and has Wi-Fi and LTE connections.
Currently, Optimus Prime has a maximum 200-degree free movement dimension in its entire body, and the hands have a maximum 27-degree movement dimension. The power used for sitting is 100 W, and the power used for walking is 500 W.
At the AI DAY scene, Optimus Prime was able to slowly walk to the stage without external assistance and did a dance move; it is reported that Optimus Prime only took the first step in April this year, and in less than half a year, it has been able to walk independently and its movement speed has been significantly improved.
In the video demonstration, Optimus Prime has been able to perform operations such as handling and watering; this is mainly due to its technical design in perception, execution, motion balance, etc.
is based on imitating the way human joints and muscles work, and Tesla has developed 6 different actuators for Optimus Prime. For example, the knee is constructed from multiple mechanical actuators with optimized joints, and the hands are similar to the human palm, with six actuators and 11 degrees of freedom on the fingers, and sensors are installed at joints and fingertips, which can pick up 20 pounds in one hand.
In terms of balance, Tesla engineers developed a new robot motion model through motion planning and control stack, plan the preset routes in advance and input them into the system, draw the next end point of the robot in the system, and then let the actuator execute and coordinate with the robot's center of gravity to maintain the balance of actions.
In terms of perception control, Optimus Prime robot uses Tesla's artificial intelligence software and sensors similar to advanced driver assistance functions. It has the same technology as the FSD beta version. It uses a camera to input data and calculates it with neural network .
Tesla Autopilot Engineering Director Milan Kovac said that experience in developing assisted driving for Tesla Autopilot, especially computer vision systems, can help them adapt to practical application scenarios.
For future research and development directions, Tesla engineers said they are developing a special battery and brake to work hard to reduce the power consumption and number of parts of Optimus Prime robots, so that the robot can run for a whole day on one charge.
In fact, Musk announced the "Tesla Robot" at AI DAY last year and said that it would completely change the company's assembly line and manufacturing business, but it was only stuck in the concept version at the time.
With the arrival of Optimus Prime humanoid robots, Musk sets up FLAG again - saying that production will reach millions as early as next year, with output expected to reach millions each, with each expected to be less than $20,000.
"It is expected that within 3 to 5 years, consumers will be able to buy Optimus Prime humanoid robots," Musk said in an interview.
It is worth mentioning that during the process of AI DAY, Musk also stepped down and posted a tweet saying, "The Optimus robot will definitely have a Catwoman version." The application scale of
FSD has reached 160,000. The first Dojo ExaPod
will be built by 2023. In addition to bringing Optimus humanoid robots, Tesla also announced the progress of FSD and Dojo supercomputers on AI Day.
Ashok Ellaswamy, head of Tesla Autopilot, said that Tesla's FSD's application in vehicles has expanded from 2,000 to 160,000 last year.
As of now, Tesla has 35 FSD software versions.According to Musk, the latest version of FSD will be launched in North America at the end of this year. Due to the impact of policy regulation and testing barriers in some countries, other markets will need to be determined according to the circumstances.
During the demonstration session, Pail Jain, the head of Tesla's autopilot motion planning technology, demonstrated the decision-making upgrade of FSD in practical applications.
For example, when turning left and entering an intersection, Tesla can calculate a variety of possible motion trajectories based on the trajectory prediction algorithm, filter out unsafe and non-compliant trajectories based on lane lines, pedestrians, road conditions and other traffic factors, and choose a track that will not make close contact with pedestrians and obstacles. Not only that, the run time of each operation of FSD has been reduced to 100 microseconds.
This is actually due to the improvement of the FSD system's huge amount of real road data learning and Tesla's data processing capabilities.
According to engineers, Tesla has developed an automatic marking technology that can process 500,000 video clips a day. Although it will eventually be reviewed by humans, its work efficiency has been greatly improved.
In addition, Tesla trains artificial intelligence software on its supercomputer and then updates to user vehicles via OTA. Not only that, Tesla also has a built-in Unreal Engine simulator for improving autopilot.
For the large amount of data collected, Tesla uses a large Nvidia GPU supercomputer and a data center to store 30PB of storage materials, but this is also difficult to meet the growing data storage needs. At the
launch conference, Tesla said it is using its own D1 chip to build a supercomputer cluster. The single training module of the Dojo supercomputer is integrated with 25 D1 blocks. It can replace the 72 GPU racks composed of 4,000 GPU with just four computing modules. The
D1 chip is a neural network training chip independently developed by Tesla. It was released on AI DAY in 2021. It is equipped with 50 billion transistors, the thermal design power consumption (TDP) reaches 400W, and the peak computing power under FP32 accuracy can reach 22.6 TFLOPS.
As a comparison, Nvidia A100 Tensor Core GPU transistors have 54.2 billion, and the thermal design power consumption is also 400W, while the peak computing power of FP32 is only 19.5TFLOPS.
According to Tesla, it plans to build the first Dojo ExaPod in PaloAlto, California, USA by 2023, with a total of 7 units.
Tesla team introduced that the construction of Dojo will gain weight and improve efficiency, and the work that originally took several months will be reduced to one week. It is not difficult to foresee that after the system is successfully built, the Dojo supercomputer is expected to become one of the most powerful supercomputers in the world.
Regarding DOJO's business model, Musk said in an interview that DOJO is not only a product for self-use, but also considers that AWS is also used for commercial use in the form of a service.
Judging from this year's AI DAY, the emergence of Optimus Prime robot undoubtedly triggered a climax of the venue. In just one year, from PPTs that exist in conceptual form to the real creation of physical robots, it is enough to see Tesla's powerful AI technology research and development capabilities.
Not only that, FLAG, which is expected to sell for around $20,000, has been mass-produced for three years and commercial use for five years, has caused a wave in the technology circle. However, given that Musk frequently releases pigeons in the application of autonomous driving technology, it remains to be verified whether robots can achieve mass production and commercial use as scheduled.
In addition, around intelligent driving, Tesla spent a lot of time explaining the technological iteration and updates when introducing the progress of FSD. It is not difficult to see that Tesla has begun to solve the long-tail problem of autonomous driving, which is also a key point related to safety.
Overall, compared with the stunning debut of the robot, the FSD part is slightly less highlighted, but from another perspective, this also indirectly reflects that it is not easy to achieve autonomous driving. Even if it is powerful like Tesla, it will have to go through a difficult climbing stage.
