Tesla humanoid robot Optimus Prime's debut did not surprise the audience, but its price as low as $20,000 will undoubtedly accelerate the commercialization of humanoid robots.
On September 30, local time, Tesla's second AI Day ( artificial intelligence nd) was held in Palo Alto, California, USA. This event is as expected by the outside world, and Tesla has made a detailed introduction to the widely-watched humanoid robots, fully autonomous driving functions, and DOJO supercomputers.
Full autonomous driving system FSD Beta will be launched worldwide by the end of this year, and its number of users will also increase from 2,000 in 2021 to 160,000 now; DOJO supercomputer is expected to become one of the most powerful supercomputers in the world. It is expected that by 2023, Tesla will build its first Dojo ExaPod in Palo Alto, California, USA - a supercomputer cluster built with D1 chips; Optimus humanoid robot is undoubtedly the protagonist of this AI. Engineers spent a lot of time explaining Optimus' logic, bionic , algorithms, etc. Tesla CEO Elon Musk 's speech and audience Q&A almost all revolve around Optimus.
Optimus Prime prototype unveiled at
As a job fair for AI talents, the entire press conference was full of obscure professional vocabulary, but this did not hinder Optimus Prime's rapid "out of the circle". Once released, the "Tesla humanoid robot prototype unveiled" became the hot search on Weibo .
Interestingly, there are many jokes on social media such as "Optimus Prime is Tie II (the humanoid robot released by Xiaomi a few days ago is Tie Ionda)" and "We owe Tie Ionda an apology" and on Twitter , many people also said that Optimus Prime is hard to see Boston's dynamics, "a light year away."
Does Optimus Prime really make netizens look forward to a year in vain?
software and hardware are highly bionic: not only look like human
just started, Tesla directly unveiled the Optimus Prime prototype. This is also the first time Optimus Prime has made a debut without any support.
"This is actually the first time a robot has stepped onto the stage without a tether," Musk said. "Robots can actually do much more than we show you. We just don't want it to fall on your face." Optimus Prime on
moves slowly, but it is worth noting that Optimus Prime's bionics are not just bionics in appearance and movement. Behind this, Tesla wants to give it a human brain through AI technology.
Tesla said that Optimus Prime's design is based on the human body and hopes that its behavior will be as close as possible to human behavior. Just as Musk insisted that autonomous driving should use visual algorithms like humans many years ago, Optimus Prime's software and hardware design also insisted on trying to "like humans".
Optimus Prime is equipped with a 2.3kWh battery pack, about 172CM tall, weighs 73kg, energy consumption 100w when sitting in a while, energy consumption 500w when walking slowly, and freedom of the entire body exceeds more than 200 degrees of freedom. Its bionic fingers are very flexible, have 27 degrees of freedom, and can walk naturally.
Judging from the data currently disclosed, Optimus Prime has become much fatter than last year's concept machine. On Tesla AI Day in August 2021, Tesla disclosed that the Optimus Prime robot is about 172CM tall and weighs about 56KG.
The next generation of Optimus Prime samples were brought to the stage
Optimus Prime iteration diagram
The video displayed by Tesla shows that at present, Optimus Prime can squat, hold boxes, grab objects, and even water flowers, and can be deployed in the Tesla factory for work and form a memory of the walking path.
In addition, Optimus Prime has applied a large amount of Tesla car development logic, such as reducing the number of parts and complexity, and collecting sensors, charging management, fusion calculations, etc. into the torso battery pack to complete; and according to the car impact test, simulate the robot fall scene and protect the robot brain, that is, the central computing platform.
In terms of Optimus Prime's walking ability, Tesla's fully autonomous driving system FSD is directly applied to Optimus Prime. Optimus Prime uses the same visual perception as a car, uses the camera to obtain data, and calculates it with the neural network .
Musk: To be a "useful robot", it costs $20,000
What specific things should Optimus Prime be used to do? Tesla did not give a clear answer, but Musk repeatedly mentioned that Optimus Prime should be made into a "useful robot."
Watering flowers, carrying heavy objects, playing football, accompanying you... In the future, where humans have needs, Optimus Prime may appear. Musk actually gave a very open answer to Tesla's functions.
Optimus Prime in Tesla factory
Not long ago, in an article published in China Netcom magazine, Musk wrote: "The initial positioning of Tesla robots was to replace people to engage in repetitive and boring and dangerous work. But the long-term goal is to make it serve thousands of households, such as cooking, mowing lawns, taking care of the elderly, etc.."
needs to be a "useful robot" among thousands of households, and the price threshold cannot be too high. Just as Tesla has reduced the cost of electric vehicles, Optimus Prime is also expected to use the scale effect to enter the homes of ordinary people.
Musk mentioned that Tesla will be committed to mass production and reduce costs. Optimus Prime will produce millions in the future, and its price is lower than Tesla cars, and its final price is expected to be less than US$20,000. The price of
made many people overjoyed. After all, the previous humanoid robots were expensive, and the price of robot dogs sold by Boston Dynamics reached US$75,000 per person.
In Optimus Prime, you can still see Tesla's ultimate pursuit of "economics".
Tesla senior mechanical design engineer and engineering team member Lizzie Miskovetz said that humans are also very efficient in some things, but at other times they are not so efficient. Although humans can maintain their life activities with a little food, they cannot stop metabolism when they are not working.
"On the robot platform, what we have to do is to minimize this and minimize idle power consumption," she said. The Tesla team also plans to peel as much complexity and weight from the robot's arms and legs. "We will minimize the number and power consumption of each component as much as possible, such as reducing sensors and wiring on the limbs."
is also Tesla's persistence in "economics" that makes the robot's life serving humans a step further. Musk mentioned that in the next 3-5 years, people may receive their own Optimus Prime. According to reports, Optimus Prime will be the first to test and apply it in Tesla's factory.
Compared to what Optimus can do, many people are more concerned about what it cannot do.
Many people are worried about whether Optimus Prime will become a "terminator" and threaten humans. Musk said that this will not happen. "When we designed Optimus Prime, we designed a lot of security barriers. A large part of the information is saved locally. You can press the local stop button or realize remote control. These security buttons are not tampered with."
For this reason, Musk said that he has not considered open source Optimus Prime at present, "may fall into the hands of bad people and use it to do some bad things."
In addition, Musk also said that in order to improve efficiency, Optimus Prime can have more than two hands, and its usage scenarios are open. Tesla will continue to gradually optimize Optimus Prime from the hardware and add accessories to it. "Mobile phones can constantly add functional modules, and Optimus Prime can also be added continuously. We hope that Optimus Prime can form an ecosystem."
Henri, professor of robotics at Arizona State University, Henri, Professor of Robotics at Arizona State University, Henri, Professor of Robotics at Arizona State University Ben Amor said Musk's $20,000 target price is a "good proposal" because the current cost of humanoid robots is as high as $100,000.
However, he also mentioned: "There are some differences between this ambition and what they have proposed." "When it comes to agility, speed, ability to walk stably, etc., there is still a lot of work to be done."
Aaron Johnson, professor of mechanical engineering at Carnegie Mellon University, commented that the demand for robots is debatable. “What is really impressive is that they have reached this level so quickly. But what is still uncertain is what the use scenarios of these millions of robots will be like."
will launch the supercomputer cluster next year ExaPod
At this AI day, Tesla also brought the latest progress of the supercomputer Dojo. The team said it plans to launch the first Dojo ExaPod in the first quarter of 2023, which is a supercomputer cluster built by Tesla using D1 chips.
Dojo's self-developed multi-chip modularized supercomputer was unveiled on last year's AI day. Initially, Dojo mainly served data annotation and training of autonomous driving systems, and now this technology is also used in robot research and development. Optimus Prime's "brain" is equipped with Dojo D1 supercomputer chip.
D1 supercomputer chip is the foundation of the Dojo system. The chip adopts a 7-nanometer manufacturing process and has a processing capacity of 1024 gigaflops, which is 102.4 billion times per second. A single chip contains 50 billion transistors , and the line length is as high as 11 miles.
1500 D1 chips can form an array, and 25 arrays can be placed on a wafer to form a training module (Training Tile), which is the core of the Dojo supercomputer.
Dojo ExaPod is linked by multiple Dojo cabinets and can be regarded as a Dojo cluster. ExaPod contains 120 training modules (Training) Tile), 3000 D1 chips. It has 13TB static random access memory capacity and 1.3TB high transmission bandwidth memory capacity, and its computing power is up to 1.1EFLOP.
ExaPod data
Tesla team introduced the performance of the Dojo cabinet in detail. A Dojo cabinet contains two layers of computing trays, each computing tray contains 6 training modules (Training Tile), with computing power of 54 PFLOPS, with 100 Power operation above kW.
6 training modules
Each layer of training module is connected to an ultra-high-speed storage system, forming 640GB of running memory, 18TB of computing bandwidth per second, and network transmission speed exceeding 1TB of network transmission speed per second.
Tesla team focuses on improving Dojo's work efficiency. With the exclusive compiler, the training delay of the same scale GPU (graphics processor) is 150 milliseconds, while Dojo only takes 5 milliseconds. By the first quarter of 2023, Dojo will get more optimizations, and the single-chip training speed may reach 4.4 times that of Nvidia A100. Tesla team introduced, Enhance Dojo efficiency by improving integration. However, high integration also brings heat dissipation problems. If the thermal expansion coefficient (CTE) is too high, the system materials will be seriously damaged.
By 2023, the Dojo single-chip training speed may be 4.4 times that of Nvidia A100. For this reason, Tesla's self-developed VRM (voltage regulation module) will be iterated within two years 1 4 versions. Currently, this module can reduce the maximum coefficient of thermal expansion (CTE) by 54%.
CTE by as much as 54%
FSD will be open globally by the end of
Since the release of Dojo supercomputer last year, the development process of Tesla FSD has continued to accelerate.
Ashok, Director of Tesla Autopilot Elliswamy mentioned that through nearly two years of testing on roads in North America, the FSD beta version of Tesla's fully autonomous driving system currently has 160,000 users, while there are only 2,000 in 2021.
Musk revealed that FSD beta will cover the world by the end of this year, and FSD Beta can already be used for road conditions in all regions around the world. However, he also mentioned that Tesla still faces some regulatory obstacles outside of North America.
At this year's shareholder meeting, Musk revealed that Tesla has made "some very important architectural improvements" to FSD, such as improving the complex problems when turning left. He said that Tesla is currently deploying FSD widely in North America. The beta version will launch a large-scale test of the fully autonomous driving capability (FSD) software this 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.
Currently, Tesla's FSD's recognition capabilities and execution decisions have been upgraded and optimized. The FSD Beta version can be automatically driven from a parking lot to another parking lot, and autonomously driven from point to point urban roads.
is also improved in the difficult unprotected left turn scenario. Through the application of neural network learning and a large amount of simulation calculations and simulations, Tesla has obtained information on identifying surrounding vehicles and pedestrians, calculated a reasonable and safe route trajectory, slowed down to a safe speed domain and passed through the gap between pedestrians and vehicles.
At the same time, just like Optimus Prime, FSD's decision-making methods are also approaching humans.
Paril Jain, head of Tesla’s Autopilot motion planning technology, demonstrates how FSD can improve in specific interactive scenarios and can make “human-like” decisions. For example, when Tesla turns left to enter an intersection, it selects a trajectory that will not come into close contact with obstacles such as pedestrians.
FSD has been improved in specific interactive scenarios
In addition, during the event's Q&A session, Musk also mentioned that he is currently driving the alpha version of FSD beta, which merges two stacks (city streets and highways).
This may be the long-awaited FSD beta V 11 in the industry. Previously, media reported that the FSD beta V11 will be Tesla's first single-stack FSD beta version. The significance is that the FSD beta V11 will be a system with all driving skills, more like a human driver. The current mainstream autonomous driving system is composed of different solutions such as ACC and NOA.
According to Musk, Tesla's goal of artificial intelligence day is still to attract talents in the field of artificial intelligence. During the event, Musk mentioned that he hopes more talents will come to Tesla to enjoy his career. "There are many companies in Silicon Valley . They look good on the surface, but they don't actually look like that. Tesla is different. If you are talented, you will get more freedom in Tesla."
