Tesla's humanoid robot development history, main achievements and technological progress

In the 1970s, Waseda University in Japan developed the world's first full-size humanoid "WABOT-1". Since then, the world has successively launched humanoid robots of different categories and applications in different scenarios. In addition, Boston Dynamics, Toyota, Honda, Xiaomi, UBTECH, etc. have all launched related products. However, the industrialization of humanoid robots is difficult, mainly due to the constraints of key technologies, high costs, and limited application scenarios. Honda, Softbank, etc. have announced the discontinuation of related humanoid robot products.

On Day 2021, Musk released Tesla's humanoid robot plan. At that time, Sla Bot was just a concept. After one year, at AI Day in September 2022, the humanoid robot Opmus was officially unveiled. Then at the shareholders' meeting in May 2023, Optimus was able to accomplish more complex tasks such as item classification. In December 2023, Optimus Gen-2 was released. Gen-2 achieved more flexible walking, could easily hold eggs and achieve more refined movements such as transfer between left and right hands. In less than three years, Optimus has achieved rapid iterative development, and the industrialization of humanoid robots may show a trend of rapid penetration from point to surface at a certain stage.

Table | Tesla Optimus upgrade status in the past two years

Source: Tesla press conference, compiled by Yifei.com

Next, I will lead you to review and summarize the development history, main achievements and technological progress of Tesla humanoid robots in the past three years. So that you can understand the reasons for the rapid iteration and upgrade of Tesla humanoid robots, and if the price does not exceed 20,000 US dollars in the future, whether the core pain points of industrialization can be broken through and whether it can drive the rapid development of the industry.

Concept

1. Tesla Bot is born

At AI Day on August 20, 2021, Tesla announced the progress of pure vision solution F, training, D1, Dojo supercomputer and other heavyweights. Finally, Tesla demonstrated the concept of its humanoid robot for the first time, named Tesla Bot. The author believes that FSD, neural network training, supercomputers, etc. seem to be supporting the automotive industry, but in fact they are the various models needed for the collection and training of humanoid robots and provide a powerful computing infrastructure, which has laid a lot of groundwork for humanoid robots.

Image | Tesla Bot

Source: Tesla AI Day

Tesla Bot is about 1.72 meters tall and weighs about 56.7kg. It integrates all the aforementioned FSD, onboard computer, Autolot camera and 40 micro-devices, and can be said to be the ultimate form of technology. In terms of technical details, the head is equipped with a camera for navigation, which is driven by an artificial neural network. There is a screen on the face for displaying and providing information, and the hands can perform movements like human hands.

2. Bottom-level FSD technology

Tesla's pure vision solution is inseparable from the multi-task learning HydraNets neural network architecture. Each Tesla car has 8 cameras surrounding the body, covering 360 degrees around it, to obtain surrounding information such as traffic lights, signs, ramps, curbs, etc., providing excellent conditions for neural network learning.

Tesla has developed the "Vector Space" technology, which combines the advantages of non-convex optimization and high dimensionality. This technology can draw a 3D bird's-eye view based on the data input from 8 cameras, forming a 4D spatial and temporal "road network" to present road information, helping vehicles grasp the driving environment and find the optimal driving path more accurately.

Source: Tesla AI Day

With massive and accurate video data, Tesla also needs to create a powerful neural network and make a special layout of the network so that the data can be integrated and re-analyzed on a general backbone network. Therefore, Tesla "started from scratch" and independently developed a training method based on neural networks.

At the same time, Tesla has also developed "scenario technology" that can be used for autonomous driving training in "edge scenarios" that are not common in reality. In simulation scenarios, Tesla can provide different environments and other parameters (obstacles, collisions, comfort, etc.), greatly improving training efficiency.

Source: Tesla AI Day

As a result, Tesla's FSD system can achieve ultra-high efficiency of 2,500 searches every 1.5 milliseconds, predicting various possible situations and finding the safest, most comfortable and fastest autonomous driving path among them.

3. Dojo Supercomputer

As the amount of data to be processed begins to grow exponentially, Tesla is also increasing the computing power for training neural networks, so the Tesla Dojo supercomputer was created. The key unit that makes up the Dojo supercomputer is the neural network training chip independently developed by Tesla - the D1 chip. The D1 chip uses a distributed structure and 7-nanometer process, with 50 billion and 354 training nodes. The internal circuit alone is 17.7 kilometers long, achieving super computing power and ultra-high bandwidth.

A single training module of the Dojo supercomputer consists of 25 D1 chips. Since each D1 chip is seamlessly connected to each other, the latency between adjacent chips is extremely low, and the training module retains bandwidth to the greatest extent. With Tesla's self-created high bandwidth and low latency, the computing power is as high as 9PFLOPs (9 quadrillion times) and the I/O bandwidth is as high as 36TB/s in a volume of less than 1 cubic foot.

Figure | Tesla FSD chip + D1 chip

Source: Tesla AI Day

So far, Tesla has tailored the most basic FSD technology and Dojo supercomputer platform for Tesla Bot, and is just waiting for the birth of Optimus. Musk also added: "There will be no shortage of labor in the future, but manual labor is only an option. Tesla Bot can perform some dangerous, repetitive and boring tasks."

Optimus Gen-1 released

1. Prototype release

One year later, on September 30, 2022, the first humanoid robot prototype "Optimus Gen-1" was officially unveiled at AI Day. The on-site demonstration showed the prototype walking and saying hello, and the video demonstration showed watering plants in the office, picking up structural parts in the factory, and recognizing renderings of surrounding objects.

Optimus Gen-1 is equipped with a 2.3kWh battery and an integrated design that can provide 52V voltage. There are 28 joints in the whole body, with more than 200 degrees of freedom. The palm design uses six actuators to complete 11 angles of free movement and can lift more than 20 pounds. Simulation modeling was carried out for the force/torque required for each joint, and 6 special motors were developed accordingly. The ball motor technology, which is the same as that of Tesla's car motor, can realize the rotation and linear motion control of each joint.

Figure | Actuator Technology

Source: Tesla AI Day

Optimus Gen-1 is equipped with the same FSD technology and Autopilot-related neural network technology as Tesla vehicles. Its brain uses the self-developed AI training Dojo D1 chip and supercomputer Dojo. After the actual application verification of the fully autonomous driving capability system, Tesla's powerful FSD technology can be directly applied to robots.

Optimus Gen-1 was finally born by equipping it with a visual camera, combining 28 joint actuators throughout the body, as well as the underlying FSD visual algorithm and Dojo supercomputer to provide computing power support.

2. First Iteration

March 2023: After five months of development, the humanoid robot is shown in the video to be able to walk independently and perform assembly tasks on another robot, with finger joints that can handle tasks such as grasping power tools, screws, and cloth covering a picture frame.

The Optimus motor's torque control and force control are more precise and sensitive; its environmental perception and memory capabilities are improved, so it can not only see the road, but also remember it; it can also perform end-to-end motion control based on human motion examples.

As of May 2023, FSD Beta has traveled nearly 200 million miles. FSD can be applied to humanoid robots. The essence is that the simulation process of autonomous driving is the same as that of robots, that is, "input-computer platform-output". The visual system recognizes the surrounding environment and transmits data to the computing platform. The platform transmits feedback to the actuator and other links to generate action. Based on FSD's technology and data accumulation, building core competitiveness, Tesla Optimus will have intelligent upgrade advantages that are different from other robot products.

3. Second Iteration

Optimus has the ability to self-calibrate its arms and legs, and using only vision and joint positions, it can accurately position its limbs in space.

Visual perception has been significantly improved, and algorithms and models have continued to improve. The robot Optimus can accurately locate its limbs in space through vision and joint position encoders, can self-calibrate its arms and legs, and can autonomously filter, classify and sort items by color, with neural networks enabling end-to-end training and learning. It can be seen that Tesla has connected the underlying models of FSD and robots, and has continuously optimized the FSD algorithm by relying on neural network models and visual technology, performing well in terms of perceiving the environment and autonomously analyzing movements. At the same time, the Tesla Dojo platform is expected to further enhance the robot's AI capabilities, and the mass production of Tesla humanoid robots is expected to accelerate.