![To prevent loss, the elevator goes directly to the safe island newspaper Liu Yadong A Source: Xinzhiyuan Xinzhiyuan Report Editor: Editorial Department [Xinzhiyuan Introduction] On Tesla AI Day, Musk proudly launched](https://cdn-dd.lujuba.top/img/loading.gif)
Prevent loss, elevator goes directly to
Safe Island reporter Liu Ya East A
Source: New Zhiyuan
9xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
Edit: Editorial Department
[New Zhiyuan Introduction] Tesla AI Day, Musk proudly launched "Optimus Prime 2.0". Have you seen a robot that is good at heart?
Tesla humanoid robot is finally here!
Just now, the second Tesla AI Day was officially launched in California . The editor who watched the full press conference at
said that this press conference was indeed very "hard core".
Musk hadn't said a few words before he came on stage, so he invited "Optimus Prime" out.
, the audience cheered.
"Optimus Prime" who has not yet prepared a shell first moves his hands as soon as he appears, and then slowly moves to the center of the stage.
You should know that this is the first time it has achieved independent walking.
Then, "Optimus Prime" raised one arm and greeted the audience below.
also raised his hands and turned on the self-satisfaction mode.
Musk said that "Optimus Prime" production is expected to reach millions, which is much cheaper than cars, and can be owned for less than 20,000 US dollars (142,320 yuan).
Musk stood up for Optimus Prime on the spot
, but this is not important. Apart from showing off his skills, "Optimus Prime" looks temporarily disappointed the public.
Optimus Prime debuted, and it costs less than US$20,000
I still remember this last year, Tesla Bot made its debut.
on the site, researchers said that the first time it was unveiled was just a concept machine. By this year's AI day, the second iteration had been achieved, and the real image was further closer to the concept image.
, and the three strong men on the scene also brought out the "Optimus Prime" that was closest to production.
I have to say that Academician Ma’s team is still very hard-working, and Academician Ma also expressed his gratitude to his engineering team.
What can this robot do? Not only can you water flowers and transport spare parts on the production line
, but you can also help you get a courier.
Next, we will enter the main topic. What technical support is behind "Optimus Prime" being able to complete this series of actions?
After two major version improvements, the current humanoid Optimus Prime can be said to be worthy of its name. It is 173 in height and has a net weight of 73 kilograms, which is similar to an adult man. In terms of power consumption control of
, the sitting state is 100W and the fast travel state is 500W. The freedom of the whole body is as high as 200, and the freedom of the hand alone is as high as 27.
The latest generation of robots (picture) In the overall body structure, orange is the actuator/execution unit, blue is the electric drive system/computing unit, equipped with a 2.3kWh battery pack, equipped with 28 structural actuators, and hands with a degree of freedom of up to 11.
Robot brain runs on Tesla SoC and has WiFi communication, audio, and security protection functions.
According to reports, the basis of Optimus Prime architecture is to apply a large number of Tesla automotive engineering ideas and development logic.
For example, based on car impact tests, simulate the fall scene of a robot to protect the "brain" central computing platform. Maybe there will be a robot's collision test in the future ...
In addition, by comparing the power systems of Tesla cars and robots, the robot will appear more complicated.
From the density of the red components below, it can be seen that the robot's power system can achieve richer functions than Tesla's autonomous driving cars.
In terms of actuator unit design, the Tesla team has also come up with 6 unique structural design solutions to meet different load needs while achieving lightweight as much as possible.
In the specific configuration of the actuator, different internal structures and bearing equipment solutions are adopted for different loads and motion fineness requirements.
Such as cross roller bearing on the left, four-point contact bearings and ball bearings on the right, etc.
Just mentioned that Tesla robots have a total of 28 joints, and each joint has multiple sets of actuators. These actuators will analyze their execution efficiency in the background to reduce invalid work.
More importantly, these actuators can achieve linear stretching under high pressure, such as lifting a piano of nearly half a ton.
Since it is " human robot ", how to make the robot's movements more like humans? The answer is to rebuild the "joint".
The actuator of Tesla's humanoid robot is inspired by the movement trajectory of human skeletons to achieve difficult movements involving multiple degrees of freedom, such as free walking, squatting, and picking things.
can be seen that in all directions, Tesla's humanoid robots can move freely.
Let’s look at the Tesla robot hand again. You should know that this is also the trailer of this AI day’s “heart-to-heart”.
It has 6 actuators, 11 degrees of freedom, supports adaptive grabbing, and can grab 20 pounds, which is especially suitable for grabbing small precision parts.
In terms of software and hardware configuration, Tesla cars and robots adopt the same visual perception. Tesla Bot's navigation system is an interesting 3D path navigation.
developers emphasized that Tesla robots’ AI technology is the same as FSD!
means that it will identify the environment itself, find what it needs, and do what it wants to do.
Tesla robot also supports indoor path planning, which not only can walk, but also walk like a human, and can also judge objects through visual recognition.
When it comes to robot walking, Tesla engineers also took us to review the "evolutionary history" of Optimus Prime walking.
took the first step from April this year to unlocking Pelvis in July, to swinging the arm in August, and to today, finally walking off the ground.
The rapid progress in the past few months, the technical core condensed behind it is not as simple as walking.
It involves the design of multiple stages such as physical self-perception, energy-efficient footwork training, balance, and coordinated movements.
For this reason, Tesla designed and planned a "motion plan" system, starting from the actual expected motion path, and generates a reference projection through a motion planner (Locomotion Planner) to generate a robot's pace motion plan.
Under the guidance of reference projection, the robot proceeds according to the pace planner.
motion simulation diagram is like the following, it would be great if it could also walk so steadily in the real world.
In fact, in the real world, the environment is complex and rich in background, and it is far from simple as the ideal environment simulated by theory.
If you are not careful, you will be off the ground with your feet and fall into a "dog gnawing on the mud" in minutes.
Regarding the process of how robots move things, the researchers said that first let someone demonstrate this action and then conduct a motion capture, let the robot train in a twin world.
Here Optimus Prime's home court has basically ended.
Tesla researchers said that in the next few months, we will work hard to make robots more flexible, achieve better balance control, and navigation, hoping to further expand Optimus Prime to help us handle more things. Although
is a bit ugly, it has indeed opened up an era.
FSD fully autonomous driving: 160,000 people use it and they say it's good to use it
Tesla FSD Beta now has 160,000 customers, while in 2021 there are only 2,000 customers.
Tesla is expected to release v.10.69.2.3 after AI Day, but no exact date has been announced.
In the past year, Tesla’s autonomous driving team has trained more than 75,000 models, which is equivalent to training one model every 8 minutes.
. Among these more than 70,000 models, the team has engineered a total of 281 models!
Master Plan
Faced with the complex road conditions in reality, Tesla needs to make a decision every 50 milliseconds.
has developed an "interactive search" framework for this team to make optimization choices.
In this way, the time of each action can be reduced from 125 milliseconds of the traditional method to almost 100 microseconds.
occupying the road to avoid identification
To model the three-dimensional world around the car, the Tesla team established this network as the Tesla team's solution, which is not currently displayed in our customer-facing visualization, what we see here is the road network output by our internal dab tool.
FSD can calculate the environment and potential collision targets on both sides of the driving route in real time, and automatically image dangerous targets, and complete the calculation within 10 milliseconds.
uses volume rendering technology to perform 3D depth reconstruction. The deep computing model is trained using a large automatic annotation dataset.
camera and target recognition
camera imaging is an important technical basis for autonomous driving technology. Even with lidar , cameras are still irreplaceable in autonomous driving.
The visual part of Tesla's FSD pathfinding neural network uses three sets of signals from the main camera, the left camera and the rear camera as input. It is processed by neural network such as Transformer, and is used as the output of the pathfinding navigation signal together with the map information, and then automatically selects the driving route through decoding and conversion by the language module.
can generate different "driving line languages" according to the encoding of the vector space to guide the vehicle to drive.
For example, what to do if you encounter a vehicle running a red light, what to do if you encounter a roadblock in front of you.
The number of FSD network parameters in the car is about 1 billion, forming a neural network with a layer of 15w or above, and the number of connections is as high as more than 37.5w. Before
automatically labeled
, Tesla's autonomous driving team relied on manual data annotations to identify and describe objects in short video clips taken by cameras and sensors on Tesla cars.
These marked fragments are used to train Tesla's neural networks and improve the driving assistance system so that Tesla cars can circulate under the supervision of drivers and automatically avoid obstacles.
Now, Tesla has developed automatic labeling technology, which can handle 500,000 fragments every day. Although
may still require humans to finalize it, at least it doesn’t need to start from 0.
solves automatic 3D annotation in three steps.
Step 1: High-precision path projection. Mark out the points, lines, and surfaces on the map.
The second step is to reconstruct the multi-path model based on the previous projection mark.
is followed by the last step, and these new routes are automatically marked.
Based on this, Tesla can improve decision-making in different environmental situations, such as night, fog and rain.
Realistic Simulation
Now, there are more than one million vehicles equipped with cameras driving on the road, and there is a simulator built in Unreal Engine to improve autonomous driving.
At least from the effect released, the reality is full!
Dojo supercomputer: 10 trillion times per second, 2023 arrives
Tesla and Dojo's goal is to build a single accelerator. One of the key steps to achieve its goals in
is its training tiles, which will be launched during AI Day 2021.
Tesla has been trying to figure out how to make its Dojo design scalable and has faced challenges in the process. One of the main purposes of the emergence of
Dojo is to be more powerful than commercially available cloud computing capabilities. Its design took into account the hardware level and should be considered for deep neural network training.
Dojo supercomputer structure key components, first including "D1 chip", providing AI computing power 362TFLOPs.
interface processor provides memory broadband and communication interface, PCIE board form. The single chip can provide 800GB/S bandwidth for training. Single chip 32GB DRAM, 900TB/s high-speed communication protocol .
is combined with 6 D1 chips, plus 20 interface processors to form a pallet, and then a cabinet Dojo Cabinet is formed by multiple pallets.
will eventually develop into a complete Dojo cluster or "ExaPod".
Dojo ExaPod, which contains 1.3TB of high-speed SRAM and 13TB of high-bandwidth DRAM, will break the computational volume of ExaFlop barrier.
Each Dojo ExaPod integrates 120 training modules, has built-in 3,000 D1 chips, has more than 1 million training nodes, and has a computing power of 1.1 EFLOP* (10 trillion floating-point operations per second ).
In addition, Tesla also demonstrated the DOJO supercomputer for the first time.
With the support of Tesla's self-developed chips, only four Dojo cabinets can replace 72 GPU racks composed of 4,000 GPUs.
Simply put, Dojo reduces the work that used to take several months to one week.
In addition, Dojo's overall integration is much higher than that of Nvidia A100 hosts. A "Dojo brick" can top multiple hosts, but the cost is lower than one. When
arrives, the maximum performance will be 4.4 times that of A100.
When Tesla introduced Dojo, it also put a small easter egg: training Stable Diffusion to generate several "Cybertrucks on Mars".
Finally, Tesla plans to build its first Exapod (1/7) by 2023, which is expected to greatly increase the output of automatic labeling. After the successful construction of the
system, the Dojo supercomputer is expected to become one of the most powerful supercomputers in the world.
I wonder which one Dojo can rank in the global supercomputer ranking?
Reference:
https://www.youtube.com/watch?v=ODSJsviD_SU
