Mengchen Zhan Shi from Aofei Temple
qubits | Official account QbitAI
annual Tesla AI DAY, the king is at the beginning!
Musk only took 3 minutes to briefly play, and humanoid robot Optimus (Optimus) appeared directly on stage.
is different from the smooth and round appearance of last year's concept PPT. This year's Optimus Prime metal and cable are exposed, just like the terminator in the movie.
The Optimus Prime shows walking steadily, turning around, and waving to say hello to the audience below. One of the videos in
also actually entered Tesla's office to water flowers and help employees carry express delivery.
was moved to the car factory and some people also carried bricks in a very good manner.
, Musk said that this is not the complete form of Optimus Prime robot, it is just a trial version of the prototype. The stage curtain of
has been opened again, and the "skinned" robot that is closer to the final product has also been moved up (it will take several weeks to walk down the ground).
As early as Tesla's first quarter earnings call this year, Musk said that "Optimus Prime's value will exceed that of the automotive business and exceed FSD."
Today he even bluntly stated that after reaching the mass production scale of millions of yuan, each robot will be much lower than that of a car, and is expected to be less than US$20,000.
Next, Musk explained the purpose of the event:
uses the coolest technology demonstration to attract the most talented engineers, mass-produce the results together, and help millions of people.
In Musk's eyes, the future of autonomous driving can improve human traffic efficiency by at least half an order of magnitude.
, and he has greater ambitions for humanoid machines, improving economic production at least two orders of magnitude. Why is
suitable for careers in Tesla?
Musk believes that as a listed company, if the public does not recognize what Tesla has done, it can use equity voting.
If I'm not suitable, I can even vote to fire me.
I can't do whatever I want, although many people think so. After the on-site recruitment of engineers was over, Musk himself is no longer the protagonist of this press conference.is a technical person in charge of each link of the robot to introduce his work results.
Optimus Prime has made progress in one year?
The Optimus Prime concept was first released in August last year. The development platform version I saw earlier was completed in February this year. The latest generation of achievements are still being improved. What remains unchanged is the pursuit of imitating humans, including height and body importance similar to humans, high freedom of joints throughout the body, especially hands, and trying to move closer to the low energy consumption of organisms.
latest version of Tesla Optimus Prime, the brain uses Tesla SoC chip, the energy relies on a 2.3-kilowatt-hour battery pack and 28 actuators (actuators) in the whole body. In addition, the hand has 11 degrees of freedom, and the five fingers can move independently.
However, this "brain" is installed on the chest, because it is safer here, and it is also more power-saving to place it in a centralized manner. There are 6 types of
actuators, 3 specifications of rotary actuators and linear actuators.
linear actuator with the highest specifications can be used to bring up an grand piano .
These actuators are assembled into different joints of the robot's body, and detailed parameter adjustments are also required.
For bipedal robots, walking is the most recognized major difficulty in the industry.
Walking is a difficult thing for humans. Millions of years of evolution allow you to automatically execute without thinking.
But for bipedal robots, the difficulty of the engineering can be seen by using a set of materials that are completely different from biological ones and relying on artificial design to imitate and approach.
Tesla engineers introduced that there are four points to pay attention to:
physical self-perception, energy-saving gait, maintaining balance and not falling, and coordinating movement throughout the body.
Tesla's method is exactly the same as the car manufacturing process, designs, simulates, and verifies .
Take the knee as an example, and the inspiration comes from humans, but it cannot completely imitate humans due to problems such as mechanical structural materials. In addition to taking steps with your knees, you should also consider the stress of squatting and picking objects. After the structure of
is designed, you can first go to the simulation environment to plan the actions.
Once you come to the complexity of the real world, you will still fail if you can't take three steps.
Through sensor measurement, introducing real-world complexity into the training environment has become the key to solving the problem.
In the end, Tesla Optimus Prime also trained repeatedly and even adapted to human interference before learning to remain stable while walking.
From taking the first slow step to learning to leave the ground with your toes, it took a total of 6 months.
and hand movement is also the same process. First, capture the movement trajectory of humans when moving boxes, and then plan in combination with the robot's own characteristics.
Tesla engineers introduced that the dual-motor version Tesla Auto only needs to consider several limited speeds and road conditions.
robots need to complete more body movements such as walking, going up steps, squatting, turning, and other tasks, from screwing screws to lifting heavy objects.
According to Musk's previous disclosure, Optimus Prime's official production will start as early as next year.
However, since it is Tesla AI DAY, in addition to robots, progress in autonomous driving and self-developed supercomputers must be avoided. Let's take a look below.
autonomous driving
autonomous driving As a signboard of Tesla, their number of FSD (Beta) users soared from 2,000 to 160,000 in one year.
is naturally quite short on AI Day. The core emphasizes that is more oriented towards complex scenarios and makes the technology deeper and more refined .
For example, turn left at an intersection without a crosswalk, and there are straight vehicles driving on the opposite road. You have to face a pedestrian pulling your child across the road. If the car passes quickly in front of the pedestrian, it will scare the other party and cause more unpredictable behaviors.
Tesla said that they added dimension analysis such as safety and comfort, through tree search algorithms and provide human demonstrations, allowing the system to find the "suitable" opportunity and route to pass, and finally, pass behind pedestrians crossing the road.
For example, for intersections in six directions, even if the car is equipped with 8 cameras, if other cars are blocked, it will be difficult to predict.
Based on various complex scenarios, Tesla announced a detailed architecture that allows various camera signals to pass through RegNets, but different cameras use different RegNets, and then enter the Special attention layer. This architecture is also suitable for scenarios such as vehicle and pedestrian behavior prediction.
visual route has been constantly questioning in the circle, and Tesla responded with a case on the spot.
They have announced a way to predict peripheral vehicles through semantic recognition. Even if the car in front picks up snowflakes on the highway to block the view, making the camera unable to "see" the car in front, the algorithm can still predict the existence of the car in front and make appropriate decisions through semantics.
In addition, in order to support model training for various complex scenarios, its full-stack development platform has been able to produce algorithm models every 8 minutes, and an array of 14,000 GPUs is formed to provide computing support. In addition, they also used Accelerated Video Library, which increased the training speed by 30%.
is on the edge, and they have also performed lightweight processing to reduce the delay of the system. 8 cameras quickly identify buildings and different pavement scenes in real time through neural network , and complete 3D scene generation faster.
refers to the fact that at the opening ceremony of the Austin Super Factory in April this year and the Q1 earnings call, Musk mentioned the self-driving taxi products without steering wheel and pedals. Many people predicted that he would talk about RoboTaxi today, but Tesla did not mention it at the scene.
chip and supercomputer
Last year, Tesla announced the D1 chip and supercomputer platform DOJO to accelerate the training of autonomous driving models, which attracted a lot of attention. In the subsequent question of
, someone questioned how they combined chips and achieved efficient compilation. This year, they came up with specific answers.
First of all, regarding the chip connection, they considered the thermal expansion coefficient (CTE) problem at the chip connection, so they simply made an external "skeleton" to put the chips into it, equipped with a high-speed memory module at the bottom, and created an "iron-lock chain array":
and then packaged the whole and put it in the cabinet:
. What is the result after one operation?
According to Tesla, their computing power on the classic project RESNET-50 beats A100:
In terms of alternatives, they took automatic labeling as an example and claimed that 4 DOJO cabinets can replace 72 GPU RACKS:
is more confident about the future, believing that not only will it beat 100 now, but it will achieve multiple surpassing by Q1 2023:
At the scene, Musk also emphasized that Tesla is considered by the outside world as an autonomous driving and car manufacturing company, but the facts are far more than that. He called Tesla a cutting-edge artificial intelligence company, and supplemented semiconductor is also a reflection of technology. Then he called more talents to join the company, drawing enough technology wishes (big) (cake).
In fact, before the event, Musk issued a preview early, saying that the event was to recruit artificial intelligence and robotics engineers/chip supercomputer engineers. In this way, all kinds of technical and muscle-showing moves also have the idea of eating twice a fish.
One More Thing
Musk himself didn't have many scenes in today's event. After stepping down, he turned to high-intensity network surfing
, for example, claiming that there will be a "Catwoman" version of Optimus Prime robot in the future.
or like the memes that netizens have. After reading so much
, finally let Optimus Prime give everyone a heart.
live playback:
https://www.youtube.com/watch?v=ODSJsviD_SU
