On the 24th local time, Tesla Robot’s official account TeslaOptimus updated the “Embodied Intelligence Debut”. In the introduction, Tesla said that the humanoid robot Optimus (Optimus Prime) can classify items autonomously.Neural network training is "completely end-to-end" - that is, video signal input and control signal output are achieved.
In the video, the Tesla robot demonstrated its capabilities of visual self-calibration, color sorting tasks, and balancing on one foot.It has many highlights in perception, brain, movement and control:
In terms of perception, through visual perception and joint position encoders, Optimus can automatically calibrate the limbs and accurately locate the spatial position of the limbs.
In terms of the brain, with pure visual technology and completely locally deployed neural networks, it can quickly adapt to the environment and complete multiple tasks.
It sorts blue and green building blocks into trays of corresponding colors. Even if someone disrupts the blocks while grabbing the blocks, Optimus can immediately adjust and adapt to the new environment and continue sorting the blocks. At the same time, it can also straighten overturned building blocks and perform new tasks such as "shuffle sorted building blocks".
In terms of motion control capabilities, Optimus can accurately grasp objects. When making movements, the robot's limbs, trunk, and finger movements are extremely flexible and close to humans. In addition, OptimusIt also performs multiple single-leg support sports stretching movements, and can maintain trunk balance while stretching.
It is worth noting that the input end of Tesla's end-to-end model is a video signal, that is, a mixed input of image and audio signals, but this demonstration shows that its recognition part of the neural network algorithm only uses visual information.
CITIC Securities believes that,It is feasible for the Tesla V12 autonomous driving algorithm system to be applied to both humanoid robots and cars, and it will help accelerate the generalization ability of the recognition algorithm. As Tesla demonstrates the possibility of an end-to-end technology route for robots, the solution is similar to its FSD operating model. The successful experience of FSD and Dojo is expected to accelerate the iteration speed of robots and expand from single tasks to multi-tasks.
From a segmental perspective, (1)On the input side, analysts pointed out that the image recognition demonstrated by Tesla this time can be achieved by both 2D and 3D cameras.The core of selection is cost control factors. Domestic manufacturers of core components such as vision lenses and cameras already have the production capacity of such products. Product supporting solutions and product cost-effectiveness are worthy of attention.
(2)On the output side, the Tesla humanoid robotThe output end is signal and attitude control. Joints, sensors, and control systems are crucial to the stability of the output end. There are many options in this field, and different manufacturers are expected to have advantages in different price segments.
(3)On the algorithm side, the end-to-end framework is able to train directly from the original input to the final output without manual feature engineering or intermediate stage processing. Although this video does not show its specific algorithm, based on its input and output information,It is recommended to focus on robot control algorithms based on reinforcement learning frameworks.
(4)On the hardware side, the video shows the static motion control stability of the Tesla robot and the high accuracy of the IMU used to control balance. Moreover, the "dual encoders" in the joint module do not have the disadvantage of strong resistance, and the encoders in dexterous hands also have high accuracy.The importance of sensing links such as IMU and encoder has become increasingly apparent.
When it comes to specific targets, Zheshang Securities is optimistic about the general trend of industrialization of humanoid robots and focuses on the manufacturing leader of core components of humanoid robots.
Electric motor: MOONS Electric, pay attention to Inovance Technology, Jiangsu Leili, INCO Co., Ltd., VEICHI Electric, etc.;
Lead screw: Hengli Hydraulic, Changsheng Bearing, Wuzhou New Year, pay attention to Best, Dingzhi Technology, Rifa Precision Machinery, Qinchuan Machine Tool, etc.;
Reducer: double ring transmission, Green Harmonic, Zhongli De, Shanghai Electromechanical, Hanyu Group, Fengli Intelligent, Han's Laser, Haozhi Electromechanical, etc.;
Sensors: Huayi Technology, Keli Sensing, Hanwei Technology, Xindong Lianke, etc.;
Controller: Huazhong CNC, Bozhong Seiko, pay attention to New Star, Eft, INVT, etc.;
Lightweight: Jinggong Technology, Zhongfu Shenying, Jilin Chemical Fiber, Jilin Carbon Valley, etc.;
Integration: Sanhua Intelligent Control, Top Group, etc.;
Complete robot machine: Eston, Boshi Co., Ltd., Xinsong Robot, etc.;
Special robot: Jingpin special equipment.