Optimus is still far from perfection, but Tesla 's determination to be a robot cannot be ignored.
"As usual, only few people believe in the potential of Optimus this time."
After the Tesla AI Day held on September 30, North American time, Musk, who was always confident, couldn't help but make fun of it.
After Musk started briefly, the Optimus first-generation prototype, which completed the development in February this year, slowly walked past the separated screen, waved to the audience, and "slided" an arc and turned around and turned back to the audience, and walked back to the screen. Although the movements of
were still stiff, Musk and his team revealed that this time on the stage was the first time that the Optimus prototype was walking normally without using a safety cable. "I was worried that it would come forward to a horse gamble."
was then the latest version of Optimus, launched by three staff members. Although the parts in the chest were wrapped, they were fixed on the bracket and did not walk upright independently. However, although it was fixed on the "shelf", it still made an action similar to "frog jumping".
The generation of prototypes displayed on Tesla AI Day | Source: Tesla
As early as last year, after announcing the creation of humanoid robots, the media and the industry have high hopes for Tesla's "Optimus" Optimus, and investment banks and analysis institutions have released relevant industry reports on a large scale. I have to say that there are not many "surprises" in the prototype that Musk showed to you today.
Some netizens even commented in the live forum: What is the difference between this and the Xiaomi robot?
Musk revealed at the press conference that the future price of Optimus will be below $20,000 - even cheaper than a Tesla electric car. What has given Musk confidence that Optimus, which is still "staggering", will mass-produce millions of units in the future and have a profound impact on the economy and human society?
The more "slim" robot
The Tesla prototype robot displayed at this press conference is 5 inches and 8 feet (about 173cm) and weighs 73kg. In terms of weight, it is still much heavier than the 57kg they said before.
After Tesla announced its various parameters last year, it was once one of the most anticipated parts in the industry. Because of the smaller weight and height ratio, it means lightweighting of internal motors and other components, as well as a high degree of integration in assembly.
Before the release of Tesla's prototype, Fu Chunjiang, head of UBL Humanoid Robot Innovation Center, once told Geek Park , "Tesla has many patents in motors. Its servo may have a better upward improvement when the volume is limited." Although
has not achieved the previous promise, it has made some progress compared to other humanoid robots in the industry now. For reference, the Altas of Boston Power weighs about 75Kg and is 1.5 meters tall. The Walker X, a domestic robot company UBL, weighs 63kg and is 130cm tall.
Tesla humanoid robot whole machine parameters | Source: Tesla
However, Tesla bot has a slimmer figure, to a certain extent, based on a more limited degree of freedom.
Previously, Musk claimed that Optimus would be equipped with 40 electromechanical actuators, that is, 40 joints that can be moved, but the prototype displayed this time only has 28 degrees of freedom as the Boston Dynamics Atlas.
Fu Chunjiang once told Geek Park, "40 are the choices that deal with normal walking and running and have the least hand operation. They are designed to build a general-purpose robot... If there are fewer ones, the freedom of the hands or arms will be relatively reduced."
Tesla technician explained at the press conference that humans have many joints that can do a lot of work well, but robots have so many freedoms, which is not very necessary. research team reduced Optimus's degree of freedom to a minimum of 28 . The reduction in components such as
motors can make Tesla's humanoid robots more "slim" in appearance, but the biggest intention of this move is more likely to be to reduce power consumption.
Last July, the bipedal robot Cassie completed a 5-kilometer outdoor long-distance race without charging halfway through, which took about 53 minutes. This is an exciting breakthrough for the robotics industry, where no robot has been able to run such a long time before.
During this 5-kilometer long run, Cassie fell twice, one because of calculations that caused the "brain" to overheat, and the other when the human operator made a mistake and turned too fast. This also means that even if the robot is operated by humans, the heat generated by the calculation process is enough to stop the machine from functioning, not to mention the high power consumption generated by a robot that automatically moves .
Optimus There is a highly integrated battery pack in the torso part | Source: Tesla
In addition to minimizing electromechanical actuators and sensors, Tesla also integrates sensors, batteries and cooling systems into the torso part to reduce the overall power consumption of the robot.
has a highly integrated battery pack with a capacity of 2.3kwh. The team researchers say that such a battery pack can support humanoid robots working for a whole day. Such integrated design also comes from Tesla's technology and experience accumulated from new energy vehicles.
Optimus' 28 degrees of freedom actually excludes the "flexible hand". The design of
is a separate chapter, which is enough to show how much Tesla pays to it. The team hopes that the robot can grab small and thin things in the future. However, this robot with 11 degrees of freedom and can be directly driven by fingers, and there are not many places to be displayed directly on the real machine.
autonomous driving, from wheels to double feet
Before designing a product, the first thing to do is to consider the scenario and landing issues. What exactly can Tesla's humanoid robot do? In the demonstration video released on site, the prototype showed three work scenarios, namely, moving boxes in the factory, delivering express delivery to company employees, watering plants, and a more detailed hand operation - moving metal strips. The operations that the robot needs to complete include identifying objects, traveling, and picking up and putting them down.
"This is the same process as we process autonomous driving data," a researcher said when talking about robots' recognition of objects.
In the development and commercialization of humanoid robots over the years, a very important restriction is the development of general-purpose AI. Even if hardware technology develops to a higher level, robots that lack software algorithm training will often be empty and human. Musk also pointed out the shortcomings of existing humanoid robots at the press conference - they are not smart and are still expensive.
Optimus Walking Progress Table｜Tesla
About Tesla’s entry into humanoid robots, the point that many people are excited about is that Tesla may take the path of “local optimization of hardware by stacking hardware” and solve the problem from general AI.
Tesla fully autonomous driving system FSD has indeed played a role in the development of humanoid robots. Musk said last week that Optimus will be controlled by artificial intelligence developed by the Autopilot team.
Optimus uses the same perception and computing unit as Tesla electric vehicles, including a self-developed chip and three automatic assisted driving cameras. It also directly grafts the computer vision neural network from the autonomous driving system . The visual neural network of
can render and simple model the environment around the car, and can also identify high-frequency and key feature points in the image stream captured by the car camera. These models and key points can help cars identify and predict the state of other objects in the environment for better navigation.
in visual navigation, path memory and planning, technology can also empower humanoid robots.
Optimus Real-time modeling of scenes and identification of key objects | Source: Tesla
"We have integrated robots into autonomous driving simulators," said a member of the team. However, they also mentioned that the environment faced by cars and robots are very different, and migration is not simple. They need to collect training data in new scenarios and train new algorithm models such as local motion.
In the factory scenario, Tesla humanoid robots have achieved good results. They revealed that the motion planning in the factory is very complete, and the robot has been deployed to the factory.
At the press conference, Optimus team members also introduced that they used the prediction model , which is used to predict whether the robot will fall in various postures, and the possible damage to various parts of the robot after falling. In the case of falls, collisions, etc., the robot will give priority to ensuring the safety of the "brain" to the greatest extent, followed by the safety of the torso battery pack.
Stability of standing and walking has always been a technical difficulty for bipedal robots, and it is also a very important factor in whether they can be commercially implemented. After all, no one wants to buy a robot that will fall from time to time when serving and delivering vegetables. The
robot prototype has a very slow walking speed, far from 8km/h as Tesla said before. However, as Fu Chunjiang introduced, autonomous driving has high requirements for real-time performance, and this framework is redundant to put on humanoid robots. Tesla's various capabilities accumulated in autonomous driving are likely to find a space to re-act as a humanoid robot.
is a robot, he is serious about
"The code you write will run on millions of humanoid robots around the world", "We welcome your participation."
is not just Musk. When introducing the technology of humanoid robots, members of Optimus team almost always call on more talents to participate in their work. Some people joked that this is a job fair like a press conference.
Tesla is recruiting staff for humanoid robots. According to Reuters , there are about 20 job openings for Tesla robots, including positions that design key components such as "actuators".
Tesla has taken serious attention to humanoid robots, and at the same time, Tesla's existing professional strength and technical capabilities may not be enough.
"Automatic driving is not as easy as people think, and to some extent, humanoid robots are the same," said Shaun Azimi, head of the Dexterous Robotics team at NASA.
For Tesla now launching humanoid robots, some industry professionals speculate that they are now training artificial intelligence and neural network for collecting data.
During Tesla's earnings call in January this year, Musk said: "I think that over time, this may be more important than the automotive business." The more important business he said is the Tesla humanoid robot. They also said before that Tesla humanoid robots may start mass production in 2023.
Judging from the prototype they released, Tesla does have mass production considerations. They look for commonalities from different motor data and reduce the types of motors, which can reduce the cost of mass production of robots. Out of the 28 motors announced by Optimus, there are only 6 different designs.
Humanoid robot actuator configuration | Source: Tesla
However, there are many doubts from colleagues. Boston Dynamics once directly started to argue: " Boston Dynamics has accumulated over ten years of accumulation in robots and their bipedal movements, but Atlas (Atlashml3) Boston Dynamics' Internet celebrity robots still have similar errors such as tripping, and the parts are also limited to commercialization. It can be seen how ridiculous it is that Musk believes that he can surpass our work within a year. "
Fuchunjiang believes that Tesla has some unique advantages in robot research and development and manufacturing:
Tesla has a large scale and is more likely to recruit more developers, and it may be faster to increase computing power and increase action diversity.
In terms of materials, Musk also engages in satellites and rockets, and they have the advantages accumulated in aerospace.
In automobile manufacturing, Tesla has greatly simplified chips and algorithms through platform-based methods, and has accumulated the ability to develop and simplify the platform.
When Sla manufactured cars, it specially built a very large stamping machine. The humanoid robot has five or six thousand parts. In the future, it may also accelerate production and manufacturing through an integrated approach. This is their advantage in manufacturing tools. Jonathan Hurst, chief technology officer of
Agility Robotics, said in an interview with Reuters: "A important criterion for measuring success is whether they (Tesla humanoid robots) make money from it." There are also doubts in the industry. Many practitioners believe that within five or even ten years, bipedal robots cannot be implemented on a large scale or enter a family.
Finally, Musk extended the time for Tesla's humanoid robot mass production to 3 to 5 years.
After really investing in the work of robots for a year, Optimus is still far from perfection, but the Tesla team has become more realistic.
