Report produced by: Southwest Securities

The following is an excerpt from the original report

------

1AI empowers robots, embodied intelligence development is timely

1.1 Robots are the best carrier for AI landing

Embodied Intelligence is the vehicle through which AI enters the physical world for interaction. According to the degree of intelligence, robots can be divided into two categories: non-intelligent robots and intelligent robots; At the same time, according to the presence or absence of carriers, artificial intelligence can also be divided into two categories: virtual AI and physical AI. The intersection of robots and artificial intelligence, that is: intelligent robots as the physical carrier of AI. General industrial robots can only be programmed to perform a series of repeated movements, all motion trajectories, positions, actions, and strength need to be set in advance, highlighting its "machine" attributes, intelligent robots can interact with the outside world, according to their own perception of the outside world, decide the way to complete the task, and can continue to learn and improve in failure, compared to the "machine" attributes of non-intelligent robots, intelligent robots are more like "people".

Robot Industry Report: AI empowers humanoid robots and pays attention to the development opportunities of core components

AI robots can be divided into five major parts. 1) The perception system, equivalent to the "five senses" of the robot, includes sensors such as sound, light, temperature, pressure, positioning, and contact, which are used to convert external environmental signals into information or data that the robot can understand; 2) The drive system, equivalent to the "muscle" of the robot, the motor drive includes the motor, reducer, encoder, etc.; 3) The end execution system, equivalent to the robot's "hand", is used to interact with the external environment; 4) Energy supply, power supply or battery; 5) The computing system and software are equivalent to the "brain" of the robot.

AI robots need to go through three levels when completing tasks. When given a task to a robot, the robot generally goes through three layers of information processing: 1) The first layer is to perceive, understand needs and environment. The robot senses its surroundings through sensors and recognizes the position of the task body within the environment. 2) The second layer is planning, disassembled into tasks and path planning, after the robot understands the task, it needs to disassemble the task into multiple steps, and execute the steps in order to achieve the goal of completing the task. 3) The third layer is execution, driving hardware to perform tasks, transforming motion planning into mechanical instructions, determining energy, momentum, speed and other parameters, and starting to execute tasks. Non-intelligent robots cannot perceive the outside world autonomously, requiring humans to calculate the motion path and set the motion parameters, while the intelligent robot can sense the outside world autonomously, and disassemble the task, design the path, and finally complete the task.

The humanoid robot track has ushered in a stage of rapid development. Since entering the 21st century, major manufacturers led by Honda and Boston Dynamics have laid out humanoid robot products, but limited by the maturity of hardware and software, the early humanoid robot development is slow. In the continuous improvement of algorithms and hardware, there are also many "new faces" on the humanoid robot track, and since 2022, Tesla, Xiaomi and other major manufacturers have successively launched their own humanoid robot products, especially Tesla with the goal of "mass production", which is expected to promote the rapid development of the entire industry.

1.2 AI large model helps the development of embodied intelligence

Major technology giants have launched AI large models, which are expected to solve the problem of versatility of humanoid robots. In 1950, Turing first proposed the concept of embodied intelligence, and in the decades since, embodied intelligence has not made much progress due to backward AI technology. In recent years, with the continuous improvement of hardware and software computing power, major companies have successively launched AI large model algorithms, including Open AI's GP T-4, Google's RT-1, PaLM-E, etc., which are expected to solve the problem of the versatility of robots and usher in the rapid development of embodied intelligence.

In December 2022, Google released the multi-task model RoboticsTransformer1 (RT-1), which greatly advances the robot's ability to summarize, summarize, and reason. RT-1 is a multi-task model that labels robot input and output actions for efficient reasoning at runtime. The trained model used a dataset covering 130,000 data strip points for more than 700 tasks, collected over 17 months using 13 robots. RT-1 can significantly improve the zero-sample generalization of new tasks, environments and objects by robots, and the success rate of robots performing tasks that have never been done before increases significantly, and the success rate of different environments even when there is interference is also increased; In addition, other robot data is used to train the model, so that the accuracy of the task performed by your own robot is improved.

In March 2023, Google and the Technical University of Berlin jointly released PaLM-E, a multimodal visual language model that can be used in robots, which uses visual data to enhance their language processing capabilities and can guide robots to complete complex tasks. PaLM-E is mainly based on Google's existing PaLM large language model, and adds the ability to perceive information, which allows the robot to really "understand" the task, and convert the seen image into understandable language text, so as to achieve "inferences" in the face of zero-sample new tasks. PaLM-E does not require preprocessing or annotation of relevant data.

In April 2023, Meta released the image segmentation model SAM. SAM can "clip" any object in any image by zero-sample generalization of unfamiliar objects and images without the need for additional training. Before the release of SAM, to accurately segment images, humans needed to manually segment them and then hand them over to robots to learn, which required a large number of experts to carry out highly specialized work, which was time-consuming and laborious. SAM allows the robot to meticulously annotate the segmented images it has learned, allowing the robot to understand what an object is, so it can generate masks for any object in any image or video, even objects that have not been seen in training.

2 AI-enabled analysis of Tesla humanoid robot Optimus

2.1 Software and hardware are continuously iteratively upgraded, and the core is to optimize cost and efficiency

Tesla's humanoid robot Optimus came out, with a core focus on optimizing cost and efficiency. In August 2021, Musk first released the Tesla Bot program, codenamed Optimus. At the Tesla AI DAY held at the end of September 2022, Musk unveiled the prototype of the humanoid robot Optimus, which weighs 73kg, sits quietly with 100W power consumption, brisk walking 500W power consumption, full body freedom of 200+, hand freedom 27, equipped with the same fully autonomous driving (FSD) brain as Tesla vehicles, and a 2.3KWh battery pack (integrated charging management, sensors, cooling system) can meet the needs of a day's work. The core of Tesla's humanoid robot is to reduce costs and energy consumption while meeting the function, that is, to reduce the number of parts and the power consumption of each component as much as possible, such as reducing the induction and wiring of the limbs. Musk said Tesla robots are expected to be delivered in 3-5 years, with millions of units produced and prices likely to be less than $20,000.

The software and hardware of the humanoid robot Optimus are constantly iteratively upgraded. In May 2023, Tesla's shareholders' meeting announced the recent progress of the humanoid robot Optimus, including scenarios such as robot walking, using vision to perceive the surrounding environment, precise arm control force not to break eggs, palm grasping flexible objects such as wiring harnesses, human demonstration training AI, robot repair robots, etc., indicating that the hardware and software performance of the humanoid robot Optimus has been further improved.

Tesla's humanoid robot Optimus is similar to "cars standing up and putting on their feet", so they can reuse a large number of car-related technologies for iterative upgrades:

1) Perception system: computer vision technology similar to that of cars can be used. According to the latest Tesla shareholders' meeting, the Optimus camera solution is a 7-camera configuration (3 in front, 1 on each side, and 1 below), which determines the position through the points observed by the camera in different directions, and projects it into the vector space, allowing the robot to perceive, recognize and understand the surrounding environment.

2) Drive system: 14 rotary linear actuators + 14 linear actuators;

3) End execution system: "dexterous hand" with a total of 12 hollow cup joints;

4) Energy supply: 2.3KWh battery pack (integrated charging management, sensor, cooling system);

5) Computing system and software: equipped with the same fully autonomous driving (FSD) brain as Tesla vehicles, the neural network and simulation training technology used by FSD are migrated to the robot, the only thing that needs to be changed is the training data set, through processing visual data, making decisions, including path planning, path memory, environmental interaction, navigation charging, etc.

2.2 AI multiple simulations to optimize the design scheme

Vehicle collision model technology migrates horizontally to protect the "brain" of the robot. In September 2022, at Tesla's AIDay, technicians demonstrated how to optimize the arrangement of parts of a humanoid robot with the help of existing AI collision models. The vehicle collision model will first record the data of each sensor during a physical collision, and then transfer the vehicle data and collision data into the system to generate a database, and then the AI will split the vehicle into more than 3,500 degrees of freedom, simulate tens of thousands of collisions at different angles and forces, and finally summarize all the collision data to find the shortcomings of the existing configuration of the vehicle and improve it. Engineers use this technology for the layout design of robot parts, placing important components away from the core damaged by collisions, and designing the shell to be more resistant to external forces, ensuring that the core components in the torso will not be affected even if the robot falls accidentally.

AI simulation models look for the best combination of cost and efficiency. Humans have more than 200 degrees of freedom, 27 degrees of freedom in the hands, and the power consumption of fast walking is only 500W, which is low power consumption and highly flexible. When designing a robot, it is necessary for the robot to reduce costs as much as possible while completing specific tasks, and find the best balance between motion capacity and low cost. Tesla designed 28 actuators (excluding hands) for the robot to complete different actions such as Pitch (rotation around the X axis), Yaw (rotation around the Z axis) and Roll (rotation around the Y axis), and then through AI simulation models and practical verification, under the goals of low power consumption, low cost and lightest weight, select the best design of a joint.

Improving the versatility of parts and simplifying 28 joints into 6 types of actuators. Tesla engineers used multiple scenarios to simulate the working state of the robot, calculated the consumption and mass of the actuator in each scenario, abstracted it into a point, solved the Pareto optimal of the entire particle cloud, and obtained a joint design that can meet a variety of use scenarios. After AI analysis and optimization, the joint selection was reused, and the actual actuator was finally reduced to 6, including 3 specifications of rotary actuators and 3 specifications of linear actuators.

2.3FSD+Dojo blessing to improve robot execution

FSD technology helps robots perceive the world. The FSD algorithm refers to the algorithm used in Tesla's FullSelf-Dri vi ng system to realize the vehicle's autonomous navigation and autonomous driving functions, allowing the vehicle to sense, make decisions and control in various traffic environments. The FSD algorithm mainly relies on neural networks and computer vision technology to extract information about roads, vehicles, pedestrians and obstacles by processing and analyzing real-time data obtained by sensors to achieve environmental recognition functions. At present, Tesla has opened up the underlying module of FSD and robot, reused FSD technology to robots, and played an important role in robot perception, decision-making and control.

Occupying the network allows the robot to "polish" its eyes. At the perception level, FSD uses the Occupancy Network to continuously detect obstacles in 3D space to estimate the position, size, and movement of obstacles. FSD algorithms can help robots perceive their surroundings and identify objects, people, and obstacles.

--- End of excerpt from the report Read more about the original report---

Report Collection Topic List X Regular Compilation and Updates by [Report Party].

(Special note: This article is derived from public information, the excerpt is for reference only and does not constitute any investment advice, please refer to the original report if you need to use it.) ）

Featured report source: Reporter

Technology / Electronics / Semiconductors /

Robot Industry Report: AI empowers humanoid robots and pays attention to the development opportunities of core components

Featured report source: Reporter