High-performance universal humanoid robot

Abstract

The high-performance universal humanoid robot comprises a main body trunk provided with a main control system, the main control system is provided with a plurality of standard communication interfaces and standard power supply interfaces, and the main body trunk is connected with an upper limb platform through a first universal flange plate interface; the main body trunk is connected with the lower limb platform through a second universal flange plate interface, and all actuators of the upper limb platform and the lower limb platform are in power supply connection with a standard power supply interface through power supply lines and are in communication connection with a standard communication interface through communication lines. The master control system is provided with an overall coordination control program based on deep reinforcement learning, model prediction control and a multi-modal large model; the method has more flexible hardware architecture and software platform, and task deployment can be quickly adjusted and optimized in a short time according to actual application requirements.

Description

Technical Field

[0001] The present invention relates to the technical field of humanoid robots, in particular to a high-performance universal humanoid robot. Background Art

[0002] Humanoid robots, as intelligent entities designed for human environments, have developed rapidly in recent years and are gradually being applied in industrial manufacturing, medical rehabilitation, service industries, logistics and warehousing, extreme environment operations, education and training, and other fields. With continuous breakthroughs in artificial intelligence, robotics, sensing technology, and actuator technology, research institutions and companies around the world have gradually realized the huge application potential and economic value of humanoid robots in various scenarios. For example, Chinese patent publication number CN119283060A describes a wheeled humanoid robot comprising a wheeled chassis running mechanism, a robot torso mechanism disposed above the wheeled chassis running mechanism, a robot head mechanism mounted on the robot torso mechanism, robot arm mechanisms mounted on both sides of the robot torso mechanism, each of the robot arm mechanisms being mounted with a gripper at one end away from the robot torso mechanism; a robot lower limb waist mechanism is disposed between the robot torso mechanism and the wheeled chassis running mechanism, and is connected to the robot lower limb waist mechanism. For example, Boston Dynamics' Atlas robot has achieved remarkable achievements in terms of movement flexibility, environmental adaptability, perception, and control. They possess a high degree of motor coordination, enabling them to perform various dynamic maneuvers in human environments, such as jumping, running, and climbing over obstacles. However, these robotic systems are still customized for specific tasks, lacking modular, scalable hardware, and multi-task adaptability. Agility Robotics' Cassie robot also demonstrated advanced technologies in gait control and dynamic balance. Cassie demonstrated excellent agility while navigating complex terrain, but the associated hardware architecture lacked standardization, preventing rapid adaptation to different task scenarios. Figure AI's Figure robot demonstrated deep learning-based dexterity and the ability to execute complex tasks, particularly in interaction and manipulation. However, these robots are mostly customized for specific applications and lack universal hardware architecture and intelligent control systems that support diverse task scenarios. Current high-performance humanoid robots face several technical challenges, including: Structural and hardware design limitations: Existing humanoid robots are often deeply customized for specific tasks, resulting in limited hardware versatility. Existing actuators and sensor modules lack standardization and vary significantly in size, form factor, mounting interface, communication bus, and control strategy. This results in lengthy design cycles for new models and hinders rapid iteration during hardware upgrades. The lack of standardized hardware interfaces and modules prevents robots from being flexibly adapted to different industries and tasks, limiting their breadth and flexibility of application.

[0003] Multimodal Perception and Control Complexity: With the continuous development of robotic perception systems, the amount of multimodal data generated, including vision, force, touch, and inertial measurements, has increased dramatically. Traditional control strategies struggle to effectively balance real-time, high-precision decision-making with robustness in dynamic scenarios. Furthermore, most current deep reinforcement learning methods focus primarily on training for fixed tasks and are unable to handle complex and changing environments. A single control strategy cannot meet the requirements for efficient, flexible, and robust robotic control, especially in complex scenarios where robots must be able to autonomously learn and handle diverse tasks. Energy Consumption and Long-Duration Operational Limitations: High-performance robots often come at the cost of high power consumption and wear and tear, severely limiting their potential for long-term continuous operation. Most current high-performance legged robots typically have a flight time of less than two hours, which is insufficient for many real-world tasks. Robots still face challenges in energy efficiency, power transmission, and thermal management, especially in long-duration operations or outdoor environments, where energy consumption directly impacts their ability to operate continuously and maintain viability. Difficulty in application expansion and deployment: Existing robots lack unified standards and a library of predefined task templates, making the deployment and adaptation of robots in different application scenarios extremely lengthy and complex. This not only increases the difficulty of robot promotion and application, but also significantly slows down the industrialization process. Summary of the Invention

[0004] This invention solves the problem of lengthy and complex deployment and adaptation of robots in different application scenarios, and proposes a high-performance universal humanoid robot with a more flexible hardware architecture and software platform, which can quickly adjust and optimize task deployment according to actual application needs in a short period of time.

[0005] In order to achieve the above objectives, the following technical solutions are proposed: A high-performance general humanoid robot includes a main body trunk equipped with a main control system, the main control system is provided with several standard communication interfaces and standard power supply interfaces, the main body trunk is connected to the upper limb platform through a first universal flange interface, the main body trunk is connected to the lower limb platform through a second universal flange interface, all actuators of the upper limb platform and the lower limb platform are connected to the standard power supply interface through power supply lines and to the standard communication interface through communication lines, and the main control system is provided with an overall coordinated control program based on deep reinforcement learning and model predictive control.

[0006] The control architecture of this invention improves task adaptability and precision. Combining deep reinforcement learning with model predictive control, it addresses the challenge of balancing real-time high precision and robustness, a challenge faced by traditional methods. The robot can adapt to dynamic scenarios and perform diverse tasks.

[0007] The modular upper and lower limb platforms of the present invention enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light load to heavy load. Users can meet different application scenarios without changing the hardware, reducing the difficulty of deployment and application.

[0008] Preferably, the main control system is electrically connected to an environment perception module for acquiring environment perception data around the robot.

[0009] The main control system of the present invention is electrically connected to an environmental perception module, and the environmental perception module is used to obtain environmental perception data around the robot, such as visual, force, tactile, inertial measurement data, etc. To this end, the environmental perception module includes at least one or more of a force perception sensor, a visual perception sensor, a tactile perception sensor, and an inertial measurement sensor IMU. The environmental perception data provides control input according to the global goals of deep reinforcement learning and model predictive control. The communication and data fusion between the environmental perception module and the main control system are used for real-time data processing and decision-making mechanisms for complex dynamic scenes. The multimodal perception of the present invention enhances environmental adaptability: by integrating visual, force, tactile and inertial measurement data, the robot has stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments.

[0010] Preferably, the environment perception module includes a visual perception sensor, a force perception sensor, a tactile perception sensor, and an inertial measurement sensor.

[0011] Preferably, the upper limb platform includes several seven-degree-of-freedom robotic arms of different load-bearing models, and the seven-degree-of-freedom robotic arms of different load-bearing models are all fixedly connected to the main body trunk through a first universal flange interface.

[0012] Taking a seven-degree-of-freedom robotic arm as an example, this invention offers three models: 5kg, 20kg, and 100kg, depending on payload requirements. These models all utilize a unified flange interface, the first universal flange interface, facilitating rapid end-of-line tool replacement and hardware expansion. Upgrades require only updating the control program parameter table, eliminating the need to replace customized hardware.

[0013] Preferably, the robotic arm is connected to the execution end via a third universal flange interface, and the execution end includes different models of dexterous hands or grippers.

[0014] The present invention further broadens the flexibility of the robot arm in switching or changing, and connects the robot arm to the execution end through a third universal flange interface. For example, different models of dexterous hands or grippers are connected to the robot arm through a unified flange interface, namely the third universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0015] Preferably, the lower limb platform includes a wheeled chassis or a footed chassis of different load-bearing models, and the wheeled chassis or the footed chassis of different load-bearing models are fixedly connected to the main body trunk through a second universal flange interface.

[0016] The lower limb platform of the present invention provides two chassis designs: foot-type chassis and wheeled chassis, supporting loads of 30Kg and 50Kg respectively. These models all use a unified flange interface, namely the second universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0017] Preferably, the wheeled chassis or footed chassis of different load-bearing models are equipped with 2-3 waist degrees of freedom.

[0018] Each chassis of the present invention can be equipped with 2-3 waist degrees of freedom to enhance the flexibility and stability of the robot in a narrow space.

[0019] Preferably, all actuators of the upper limb platform and the lower limb platform include at least a motor, an encoder and a driver. The drivers all use the EtherCAT communication protocol to communicate with the standard communication interface through a communication line, and the motors all use a unified power supply line to connect to the standard power supply interface for power supply.

[0020] The robot joint module of the present invention is the core of the robot execution unit and includes the following components: a motor, which provides a power source; a reducer or a screw mechanism, which is used to increase the torque; an encoder, which accurately measures position and speed; and a driver, which controls the operation of the motor.

[0021] Communication: All actuators utilize the EtherCAT communication protocol, which offers high real-time performance and strong anti-interference capabilities, supporting multi-node synchronization. To simplify wiring and improve system reliability, the drives are positioned close to the motors, reducing the cable length from the drives to the motors. Only two main cables are required: a 48V power supply and an EtherCAT communication cable to the main control system. The modular design supports standardized interfaces for actuators ranging from small (5Nm) to large (500Nm), ensuring scalability and maintainability.

[0022] Circuit design advantages: Through standardized interface design, different actuators can be plug-and-play at the hardware level, reducing development and maintenance costs and making hardware upgrades easier.

[0023] The modular design of this invention enhances hardware versatility and scalability: Through standardized interface design, execution units and hardware modules can be quickly adapted and replaced to meet different task requirements, significantly reducing development cycles and maintenance costs. Furthermore, the adoption of the EtherCAT protocol enhances the system's real-time performance and stability.

[0024] Preferably, the overall coordinated control program controls the upper limb platform including end position control and adaptive adjustment based on force feedback.

[0025] The overall coordinated control program of the present invention combines deep reinforcement learning (DRL), model predictive control (MPC) and whole-body control (WBC), multimodal large models (Visual Language Action models, VLA) and world models to achieve a high degree of unity of multimodal perception and control. The specific implementation of the control of the upper limb platform is as follows: taking the robotic arm as an example, the end position control is first performed, and the end trajectory of the robotic arm is optimized through MPC, and multiple redundant degrees of freedom are combined to maintain its balance; then force feedback and adaptive adjustment are performed, and the environmental perception data of the environmental perception module are integrated to adjust the operating force and direction of the robotic arm in real time.

[0026] Preferably, the overall coordination control program controls the lower limb platform by including offline training of a gait generation model and online optimization and adjustment of step length and foothold position.

[0027] The specific implementation of the overall coordinated control program of the present invention for the control of the lower limb platform is as follows: gait control is performed, taking a footed chassis robot as an example: first, deep reinforcement learning is performed, and the gait generation model is trained offline to generate the desired foot trajectory and control input based on the target direction and terrain information; then model predictive control is performed, and joint control instructions are optimized online in real time to ensure the stability and energy efficiency of the gait, and dynamically adjust the step length and foothold position.

[0028] In the final overall coordination, the main control system receives environmental perception data and provides control input based on the global goals predicted by deep learning; the global degree of freedom allocation is optimized in real time through MPC, taking into account both operational accuracy and dynamic balance.

[0029] The energy efficiency optimization of the present invention extends the working time: the present invention reduces energy consumption and improves the robot's ability to work continuously by optimizing gait and power transmission strategies, thus solving the problem of insufficient endurance of existing robots.

[0030] The beneficial effects of the present invention are: Modular design improves hardware versatility and scalability: Through standardized interface design, this invention enables rapid adaptation and replacement of execution units and hardware modules for different task requirements, significantly reducing development cycles and maintenance costs. Furthermore, the adoption of the EtherCAT protocol enhances the system's real-time performance and stability.

[0031] The control architecture improves task adaptability and accuracy: Combining deep reinforcement learning, model predictive control, and a multimodal large-scale control strategy, it addresses the challenge of balancing real-time high accuracy and robustness, a challenge faced by traditional methods. The robot is able to adapt to dynamic scenarios and perform diverse tasks.

[0032] Modular upper and lower limb platforms enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light to heavy loads. Users can meet different application scenarios without changing hardware, reducing the difficulty of deployment and application.

[0033] Energy efficiency optimization extends working time: This invention reduces energy consumption and improves the robot's ability to work continuously by optimizing gait and power transmission strategies, thus solving the problem of insufficient endurance of existing robots.

[0034] Multimodal perception enhances environmental adaptability: by integrating visual, force, tactile, and inertial measurement data, robots have stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a simplified diagram of the system structure of the present invention.

[0036] Figure 2 It is a schematic diagram of the plug-and-play of the actuator of the present invention.

[0037] Figure 3 It is a flow chart of the present invention for dynamically adjusting upper and lower limbs.

[0038] Figure 4 It is a schematic diagram of the dynamic adjustment of the lower limb movement pattern of the present invention.

[0039] Figure 5 This is a schematic diagram of the control strategy of the present invention that combines deep reinforcement learning and model predictive control. DETAILED DESCRIPTION

[0040] Example 1: This embodiment proposes a high-performance universal humanoid robot. Figure 1, including a main control system, a main trunk, an upper limb platform, and a lower limb platform. The main control system is set on the main trunk. The main trunk and the upper limb platform are connected through a first universal flange interface, and the main trunk and the lower limb platform are connected through a second universal flange interface. The main control system is provided with several standard power supply interfaces and standard communication interfaces. All actuators of the lower limb platform and the upper limb platform are connected for communication through communication lines and standard communication interfaces, and for power supply through power lines and standard power supply interfaces. In order to adapt to different scenarios, the main control system is provided with an overall coordination control program, which is formulated based on deep reinforcement learning and model predictive control.

[0041] The control architecture of this invention improves task adaptability and precision. Combining deep reinforcement learning with model predictive control, it addresses the challenge of balancing real-time high precision and robustness, a challenge faced by traditional methods. The robot can adapt to dynamic scenarios and perform diverse tasks.

[0042] The modular upper and lower limb platforms of the present invention enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light load to heavy load. Users can meet different application scenarios without changing the hardware, reducing the difficulty of deployment and application.

[0043] The main control system of the present invention is electrically connected to an environmental perception module, and the environmental perception module is used to obtain environmental perception data around the robot, such as visual, force, tactile, inertial measurement data, etc. To this end, the environmental perception module includes at least one or more of a force perception sensor, a visual perception sensor, a tactile perception sensor, and an inertial measurement sensor IMU. The environmental perception data provides control input according to the global goals of deep reinforcement learning and model predictive control. The communication and data fusion between the environmental perception module and the main control system are used for real-time data processing and decision-making mechanisms for complex dynamic scenes. The multimodal perception of the present invention enhances environmental adaptability: by integrating visual, force, tactile and inertial measurement data, the robot has stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments.

[0044] Taking a seven-degree-of-freedom robotic arm as an example, this invention offers three models: 5kg, 20kg, and 100kg, depending on payload requirements. These models all utilize a unified flange interface, the first universal flange interface, facilitating rapid end-of-line tool replacement and hardware expansion. Upgrades require only updating the control program parameter table, eliminating the need to replace customized hardware.

[0045] The lower limb platform of the present invention provides two chassis designs: foot-type chassis and wheeled chassis, supporting loads of 30Kg and 50Kg respectively. These models all use a unified flange interface, namely the second universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0046] Each chassis of the present invention can be equipped with 2-3 waist degrees of freedom to enhance the flexibility and stability of the robot in a narrow space.

[0047] Example 2: This embodiment optimizes the overall coordination control program based on embodiment 1 and proposes a high-performance universal humanoid robot. Figure 1 , including a main control system, a main trunk, an upper limb platform, and a lower limb platform. The main control system is set on the main trunk. The main trunk and the upper limb platform are connected through a first universal flange interface, and the main trunk and the lower limb platform are connected through a second universal flange interface. The main control system is provided with several standard power supply interfaces and standard communication interfaces. All actuators of the lower limb platform and the upper limb platform are connected for communication through communication lines and standard communication interfaces, and for power supply through power lines and standard power supply interfaces. In order to adapt to different scenarios, the main control system is provided with an overall coordination control program, which is formulated based on deep reinforcement learning and model predictive control.

[0048] The control architecture of this invention improves task adaptability and precision. Combining deep reinforcement learning, model predictive control, and a multimodal large-scale control strategy, it addresses the challenge of balancing real-time high accuracy and robustness, a challenge faced by traditional approaches. The robot is able to adapt to dynamic scenarios and perform diverse tasks.

[0049] The modular upper and lower limb platforms of the present invention enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light load to heavy load. Users can meet different application scenarios without changing the hardware, reducing the difficulty of deployment and application.

[0050] The main control system of the present invention is electrically connected to an environmental perception module, and the environmental perception module is used to obtain environmental perception data around the robot, such as visual, force, tactile, inertial measurement data, etc. To this end, the environmental perception module includes at least one or more of a force perception sensor, a visual perception sensor, a tactile perception sensor, and an inertial measurement sensor IMU. The environmental perception data provides control input according to the global goals of deep reinforcement learning and model predictive control. The communication and data fusion between the environmental perception module and the main control system are used for real-time data processing and decision-making mechanisms for complex dynamic scenes. The multimodal perception of the present invention enhances environmental adaptability: by integrating visual, force, tactile and inertial measurement data, the robot has stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments.

[0051] Taking a seven-degree-of-freedom robotic arm as an example, this invention offers three models: 5kg, 20kg, and 100kg, depending on payload requirements. These models all utilize a unified flange interface, the first universal flange interface, facilitating rapid end-of-line tool replacement and hardware expansion. Upgrades require only updating the control program parameter table, eliminating the need to replace customized hardware.

[0052] The lower limb platform of the present invention provides two chassis designs: foot-type chassis and wheeled chassis, supporting loads of 30Kg and 50Kg respectively. These models all use a unified flange interface, namely the second universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0053] Each chassis of the present invention can be equipped with 2-3 waist degrees of freedom to enhance the flexibility and stability of the robot in a narrow space.

[0054] The overall coordinated control program of the present invention combines deep reinforcement learning (RL) and model predictive control (MPC) to achieve a high degree of unity between multimodal perception and control. The specific implementation of the control of the upper limb platform is as follows: Taking the robotic arm as an example, the end position control is first performed, and the end trajectory of the robotic arm is optimized through MPC, and multiple redundant degrees of freedom are combined to maintain its balance; then force feedback and adaptive adjustment are performed, and the environmental perception data of the environmental perception module are integrated to adjust the operating force and direction of the robotic arm in real time.

[0055] The specific implementation of the overall coordinated control program of the present invention for the control of the lower limb platform is as follows: gait control is performed, taking a footed chassis robot as an example: first, deep reinforcement learning is performed, and the gait generation model is trained offline to generate the desired foot trajectory and control input based on the target direction and terrain information; then model predictive control is performed, and joint control instructions are optimized online in real time to ensure the stability and energy efficiency of the gait, and dynamically adjust the step length and foothold position.

[0056] In the final overall coordination, the main control system receives environmental perception data and provides control input based on the global goals predicted by deep learning; the global degree of freedom allocation is optimized in real time through MPC, taking into account both operational accuracy and dynamic balance.

[0057] The energy efficiency optimization of the present invention extends the working time: the present invention reduces energy consumption and improves the robot's ability to work continuously by optimizing gait and power transmission strategies, thus solving the problem of insufficient endurance of existing robots.

[0058] Example 3: This embodiment optimizes the actuator based on embodiment 2 and proposes a high-performance universal humanoid robot. Figure 1 , including a main control system, a main trunk, an upper limb platform, and a lower limb platform. The main control system is set on the main trunk. The main trunk and the upper limb platform are connected through a first universal flange interface, and the main trunk and the lower limb platform are connected through a second universal flange interface. The main control system is provided with several standard power supply interfaces and standard communication interfaces. All actuators of the lower limb platform and the upper limb platform are connected for communication through communication lines and standard communication interfaces, and for power supply through power lines and standard power supply interfaces. In order to adapt to different scenarios, the main control system is provided with an overall coordination control program, which is formulated based on deep reinforcement learning and model predictive control.

[0059] The robot joint module of the present invention is the core of the robot execution unit. Figure 2 , includes the following components: motor, which provides power source; reducer or screw mechanism, which is used to increase torque; encoder, which accurately measures position and speed; and driver, which controls motor operation.

[0060] Communication method: All actuators use the EtherCAT communication protocol with high real-time performance and strong anti-interference ability, and support multi-node synchronization. In order to simplify wiring and improve system reliability, the drive is placed close to the motor to reduce the cable length from the drive to the motor. Only two main cables are required, namely the 48V power supply line and the EtherCAT communication line connected to the main control system. The modular design supports standardized interfaces from small actuators (5Nm) to large actuators (500Nm), ensuring scalability and maintainability. The power supply system is connected to the motor drive, and the drive then powers the motor. The power supply interface and communication interface of the main control system are only connected to the first drive of the upper limb platform and the lower limb platform, and the drive is connected to the power supply line and communication line of each drive in turn.

[0061] Circuit design advantages: Through standardized interface design, different actuators can be plug-and-play at the hardware level, reducing development and maintenance costs and making hardware upgrades easier.

[0062] The modular design of this invention enhances hardware versatility and scalability: Through standardized interface design, execution units and hardware modules can be quickly adapted and replaced to meet different task requirements, significantly reducing development cycles and maintenance costs. Furthermore, the adoption of the EtherCAT protocol enhances the system's real-time performance and stability.

[0063] The control architecture of this invention improves task adaptability and precision. Combining deep reinforcement learning, model predictive control, and a multimodal large-scale control strategy, it addresses the challenge of balancing real-time high accuracy and robustness, a challenge faced by traditional approaches. The robot is able to adapt to dynamic scenarios and perform diverse tasks.

[0064] The modular upper and lower limb platforms of the present invention enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light load to heavy load. Users can meet different application scenarios without changing the hardware, reducing the difficulty of deployment and application.

[0065] The main control system of the present invention is electrically connected to an environmental perception module, and the environmental perception module is used to obtain environmental perception data around the robot, such as visual, force, tactile, inertial measurement data, etc. To this end, the environmental perception module includes at least one or more of a force perception sensor, a visual perception sensor, a tactile perception sensor, and an inertial measurement sensor IMU. The environmental perception data provides control input according to the global goals of deep reinforcement learning and model predictive control. The communication and data fusion between the environmental perception module and the main control system are used for real-time data processing and decision-making mechanisms for complex dynamic scenes. The multimodal perception of the present invention enhances environmental adaptability: by integrating visual, force, tactile and inertial measurement data, the robot has stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments.

[0066] Taking a seven-degree-of-freedom robotic arm as an example, this invention offers three models: 5kg, 20kg, and 100kg, depending on payload requirements. These models all utilize a unified flange interface, the first universal flange interface, facilitating rapid end-of-line tool replacement and hardware expansion. Upgrades require only updating the control program parameter table, eliminating the need to replace customized hardware.

[0067] The lower limb platform of the present invention provides two chassis designs: foot-type chassis and wheeled chassis, supporting loads of 30Kg and 50Kg respectively. These models all use a unified flange interface, namely the second universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0068] Each chassis of the present invention can be equipped with 2-3 waist degrees of freedom to enhance the flexibility and stability of the robot in a narrow space.

[0069] The overall coordinated control program of the present invention combines deep reinforcement learning (RL) and model predictive control (MPC) to achieve a high degree of unity between multimodal perception and control. The specific implementation of the control of the upper limb platform is as follows: Taking the robotic arm as an example, the end position control is first performed, and the end trajectory of the robotic arm is optimized through MPC, and multiple redundant degrees of freedom are combined to maintain its balance; then force feedback and adaptive adjustment are performed, and the environmental perception data of the environmental perception module are integrated to adjust the operating force and direction of the robotic arm in real time.

[0070] The specific implementation of the overall coordinated control program of the present invention for the control of the lower limb platform is as follows: gait control is performed, taking a footed chassis robot as an example: first, deep reinforcement learning is performed, and the gait generation model is trained offline to generate the desired foot trajectory and control input based on the target direction and terrain information; then model predictive control is performed, and joint control instructions are optimized online in real time to ensure the stability and energy efficiency of the gait, and dynamically adjust the step length and foothold position.

[0071] In the final overall coordination, the main control system receives environmental perception data and provides control input based on the global goals predicted by deep learning; the global degree of freedom allocation is optimized in real time through MPC, taking into account both operational accuracy and dynamic balance.

[0072] The energy efficiency optimization of the present invention extends the working time: the present invention reduces energy consumption and improves the robot's ability to work continuously by optimizing gait and power transmission strategies, thus solving the problem of insufficient endurance of existing robots.

[0073] Example 4: This embodiment optimizes the robotic arm based on embodiment 3 and proposes a high-performance universal humanoid robot. Figure 1 , including a main control system, a main trunk, an upper limb platform, and a lower limb platform. The main control system is set on the main trunk. The main trunk and the upper limb platform are connected through a first universal flange interface, and the main trunk and the lower limb platform are connected through a second universal flange interface. The main control system is provided with several standard power supply interfaces and standard communication interfaces. All actuators of the lower limb platform and the upper limb platform are connected for communication through communication lines and standard communication interfaces, and for power supply through power lines and standard power supply interfaces. In order to adapt to different scenarios, the main control system is provided with an overall coordination control program, which is formulated based on deep reinforcement learning and model predictive control.

[0074] The robot joint module of the present invention is the core of the robot execution unit. Figure 2 , includes the following components: motor, which provides power source; reducer or screw mechanism, which is used to increase torque; encoder, which accurately measures position and speed; and driver, which controls motor operation.

[0075] Communication method: All actuators use the EtherCAT communication protocol with high real-time performance and strong anti-interference ability, and support multi-node synchronization. In order to simplify wiring and improve system reliability, the drive is placed close to the motor to reduce the cable length from the drive to the motor. Only two main cables are required, namely the 48V power supply line and the EtherCAT communication line connected to the main control system. The modular design supports standardized interfaces from small actuators (5Nm) to large actuators (500Nm), ensuring scalability and maintainability. The power supply system is connected to the motor drive, and the drive then powers the motor. The power supply interface and communication interface of the main control system are only connected to the first drive of the upper limb platform and the lower limb platform, and the drive is connected to the power supply line and communication line of each drive in turn.

[0076] Circuit design advantages: Through standardized interface design, different actuators can be plug-and-play at the hardware level, reducing development and maintenance costs and making hardware upgrades easier.

[0077] The modular design of this invention enhances hardware versatility and scalability: Through standardized interface design, execution units and hardware modules can be quickly adapted and replaced to meet different task requirements, significantly reducing development cycles and maintenance costs. Furthermore, the adoption of the EtherCAT protocol enhances the system's real-time performance and stability.

[0078] The control architecture of this invention improves task adaptability and precision. Combining deep reinforcement learning, model predictive control, and a multimodal large-scale control strategy, it addresses the challenge of balancing real-time high accuracy and robustness, a challenge faced by traditional approaches. The robot is able to adapt to dynamic scenarios and perform diverse tasks.

[0079] The modular upper and lower limb platforms of the present invention enable rapid task switching: The modular design of the upper and lower limb platforms provides flexible configuration capabilities from light load to heavy load. Users can meet different application scenarios without changing the hardware, reducing the difficulty of deployment and application.

[0080] The main control system of the present invention is electrically connected to an environmental perception module, and the environmental perception module is used to obtain environmental perception data around the robot, such as visual, force, tactile, inertial measurement data, etc. To this end, the environmental perception module includes at least one or more of a force perception sensor, a visual perception sensor, a tactile perception sensor, and an inertial measurement sensor IMU. The environmental perception data provides control input according to the global goals of deep reinforcement learning and model predictive control. The communication and data fusion between the environmental perception module and the main control system are used for real-time data processing and decision-making mechanisms for complex dynamic scenes. The multimodal perception of the present invention enhances environmental adaptability: by integrating visual, force, tactile and inertial measurement data, the robot has stronger environmental understanding and dynamic decision-making capabilities, and can cope with complex and uncertain working environments.

[0081] Taking a seven-degree-of-freedom robotic arm as an example, this invention offers three models: 5kg, 20kg, and 100kg, depending on payload requirements. These models all utilize a unified flange interface, the first universal flange interface, facilitating rapid end-of-line tool replacement and hardware expansion. Upgrades require only updating the control program parameter table, eliminating the need to replace customized hardware.

[0082] The present invention further broadens the flexibility of the robot arm in switching or changing, and connects the robot arm to the execution end through a third universal flange interface. For example, different models of dexterous hands or grippers are connected to the robot arm through a unified flange interface, namely the third universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0083] The lower limb platform of the present invention provides two chassis designs: foot-type chassis and wheeled chassis, supporting loads of 30Kg and 50Kg respectively. These models all use a unified flange interface, namely the second universal flange interface, which facilitates the rapid replacement of end tools and hardware expansion. When upgrading, only the parameter table of the control program needs to be updated, and there is no need to replace the customized hardware structure.

[0084] Each chassis of the present invention can be equipped with 2-3 waist degrees of freedom to enhance the flexibility and stability of the robot in a narrow space.

[0085] refer to Figure 3 , focusing on the methods and processes by which the robot can dynamically adjust its upper limbs (from grippers to dexterous hands) and lower limbs (such as extending its body to complete high-altitude tasks) when task requirements change: Scenario task: The robot is required to take precision instruments (such as test tubes) from the desktop and place them on a high shelf.

[0086] Process demonstration: Step 1: Task instructions: The instructions are clear: when moving precision instruments, the height must be beyond the normal reach.

[0087] Step 2: Scenario Analysis: Determine the task requirements: fine operation + high placement.

[0088] Step 3: Select sports mode: The basic movement pattern is determined: grab + lift + place.

[0089] Step 4: Hardware configuration evaluation: Upper limb requirements: Fine manipulation, which cannot be accomplished by the current gripper and needs to be replaced with a dexterous hand.

[0090] Lower limb requirements: Insufficient height, limb length needs to be increased to reach shelf height.

[0091] Step 5: Adaptive hardware adjustment: Activate the hand module: Replace with a dexterous hand (which can automatically switch modules or adjust flexibility).

[0092] Activate the lower limb module: extend the lower limb joints, increase the number of trunk segments or raise the main trunk.

[0093] Step 6: Self-calibration and detection: Calibration of gripping force and feedback sensors of dexterous hands.

[0094] Balance stability calibration after limb extension.

[0095] Step 7: Action path planning: Replan the grasping path and height lifting path according to the adjusted posture.

[0096] Step 8: Execute the task: Dexterous hands gently grasp the test tube, lift it safely and stably to the high shelf, and place it precisely.

[0097] Step 9: Feedback and Evaluation: Check placement accuracy, robot posture balance, and execution efficiency.

[0098] Step 10: Complete the task or make adjustments: If the task is completed perfectly, the robot returns to the standby state; if there are problems with accuracy or posture, fine-tuning is performed again and feedback is given to the adjustment step.

[0099] Robot adaptive hardware module description: Upper limbs: equipped with quick-change devices or self-adjustable clamps: Standard gripper (strong gripping force, suitable for handling tasks).

[0100] Dexterous hands (high-precision multi-finger grasping to achieve fine manipulation).

[0101] Lower limbs: telescopic joints or increased number of joints, height adjustment: Retractable legs with hydraulic or electric extension joints.

[0102] Modular torso extension mechanism allows for quick adjustment of torso height.

[0103] Intelligent control: It has autonomous task recognition, adaptive hardware selection, and fast motion path reconstruction algorithm.

[0104] Through the above demonstration examples, this embodiment demonstrates how a high-performance general humanoid robot can autonomously analyze and quickly adjust its hardware configuration and action strategy after receiving different task instructions, thereby achieving the switching from grippers to dexterous hands, as well as the adaptive functions of lower limb extension or body elevation, giving the robot more powerful task generalization capabilities.

[0105] refer to Figure 4 When the environment changes, the robot can adaptively adjust its proportions from stairs to flat ground, and adaptively from foot-based to wheel-based. Specific examples are as follows: Scenario tasks: The robot needs to move from a staircase environment to a flat and wide area, and automatically switch its motion mode from legged (bipedal or quadrupedal) to wheeled motion, see Table 1.

[0106] Table 1 Analysis of advantages of lower limb platform in specific application scenarios Specific demonstration of the process: Step 1: Receive environmental perception data: LiDAR, visual sensors, IMU, etc. perceive the current environment characteristics.

[0107] Step 2: Environmental type analysis and identification: Current environment: The stairs are about to enter the flat ground environment.

[0108] Step 3: Exercise mode adaptability assessment: Foot type: Excellent in stair environments, but not very efficient in moving on flat ground.

[0109] Wheeled: Efficient on flat ground, but not suitable for stairs.

[0110] Step 4: Determine whether to adjust the exercise mode: Adjustment required: Switching from foot-based to wheel-based.

[0111] Step 5: Startup mode switching and adjustment: The wheeled mechanism at the end of the robot's lower limbs unfolds and locks into place.

[0112] The foot mechanism folds or lifts to complete the mode switch.

[0113] Step 6: Perform robot hardware and morphology adjustments: The foot structure is stowed towards the legs or fuselage or lifted off the ground.

[0114] The wheeled unit at the bottom extends outward to the ground and maintains firm contact.

[0115] Step 7: Self-calibration and detection after adjustment: Wheel mode motion calibration (tire pressure, bearings, drive motor status check).

[0116] Overall stability, sensor feedback calibration and self-test.

[0117] Step 8: Update the motion path planning algorithm: Switch to wheeled driving path planning and obstacle avoidance algorithm.

[0118] Recalculate the optimal path to accommodate wheeled mobility.

[0119] Step 9: Execute the movement according to the new mode: It moves at high speed in wheeled mode on flat areas, ensuring efficient movement.

[0120] Step 10: Real-time feedback monitoring: Continuously monitor the robot status to ensure movement safety and path accuracy.

[0121] Details of the robot's adaptive hardware module: Upper limbs and torso: Stability automatic balancing mechanism ensures a stable center of gravity when switching modes.

[0122] Lower limbs: Composite wheel-leg mechanism (Wheel-Leg): Foot mode: foot-end support, highly adaptable (stairs, uneven surfaces).

[0123] Wheeled mode: The built-in wheels extend to allow for fast and smooth ground movement.

[0124] Use hydraulic, electric or pneumatic modules for fast switching.

[0125] Sensing and control system: multi-sensor fusion environment recognition (vision / radar / Li DAR).

[0126] Real-time motion control and automatic path planning algorithm: intelligent feedback closed-loop control, rapid self-detection and adjustment.

[0127] Through the detailed examples above, this embodiment demonstrates the robot's adaptive transition mechanism from foot-based to wheeled locomotion in response to changing environments. The robot can autonomously identify its environment and automatically adjust its form and motion pattern to ensure efficient, safe, and stable completion of tasks in diverse environments. This highly intelligent, adaptive design significantly enhances the general-purpose humanoid robot's environmental adaptability and task execution efficiency.

[0128] The overall coordinated control program of the present invention combines deep reinforcement learning (RL) and model predictive control (MPC) to achieve a high degree of unity between multimodal perception and control. The specific implementation of the control of the upper limb platform is as follows: Taking the robotic arm as an example, the end position control is first performed, and the end trajectory of the robotic arm is optimized through MPC, and multiple redundant degrees of freedom are combined to maintain its balance; then force feedback and adaptive adjustment are performed, and the environmental perception data of the environmental perception module are integrated to adjust the operating force and direction of the robotic arm in real time.

[0129] The specific implementation of the overall coordinated control program of the present invention for the control of the lower limb platform is as follows: gait control is performed, taking a footed chassis robot as an example: first, deep reinforcement learning is performed, and the gait generation model is trained offline to generate the desired foot trajectory and control input based on the target direction and terrain information; then model predictive control is performed, and joint control instructions are optimized online in real time to ensure the stability and energy efficiency of the gait, and dynamically adjust the step length and foothold position.

[0130] In the final overall coordination, the main control system receives environmental perception data and provides control input based on the global goals predicted by deep learning; the global degree of freedom allocation is optimized in real time through MPC, taking into account both operational accuracy and dynamic balance.

[0131] The energy efficiency optimization of the present invention extends the working time: the present invention reduces energy consumption and improves the robot's ability to work continuously by optimizing gait and power transmission strategies, thus solving the problem of insufficient endurance of existing robots.

[0132] refer to Figure 5Under the management of artificial intelligence models (large models or MoE expert models), the main control system achieves high performance, adaptability, security, reliability and human-like intelligence, thus achieving universal application. Specific examples are as follows: Scenario task: The robot needs to autonomously complete the "kitchen cleaning" task in a human environment, demonstrating high performance, adaptability, safety, reliability and human-like intelligence.

[0133] Specific demonstration of the process: Step 1: Accept mission instructions and objectives: Accept natural language commands: "Please clean up the kitchen and put the dishes back where they belong." Step 2: Multimodal environment perception data collection: Use vision (RGB-D camera, lidar), tactile sensors, and sound to detect environmental conditions and confirm the kitchen layout, tableware location, stains, or garbage location.

[0134] Step 3: Large model / expert model hybrid analysis: Large model: Understands the semantics of the task and establishes an overall plan for the cleaning task; identifies the kitchen environment and locates objects such as tableware, garbage, and stains.

[0135] MoE expert model: Accurately analyzes specific action requirements such as grasping points and wiping areas.

[0136] Step 4: The large model autonomously generates behavioral strategies: Based on human common sense and experience, it can autonomously generate an overall task plan: clean up table garbage; wipe the table; wash the tableware; and put the tableware back in place.

[0137] Step 5: Model of Experts (MoE) fine motion planning and optimization: The expert model carefully plans each subtask: the action path for grabbing tableware; controlling the force and trajectory when wiping the table; and adjusting the body posture to ensure safe and reliable actions.

[0138] Step 6: Real-time action execution and control output: The robot performs real-time motion control: grabbing trash and putting it in the trash can; using its dexterous hands and tools to wipe away stains; and precisely washing dishes and placing them in the corresponding cabinets.

[0139] Step 7: Multi-dimensional feedback real-time monitoring: Real-time monitoring of robot status: safety monitoring: collisions, falls, and abnormal situations; performance monitoring: task efficiency and accuracy; adaptability monitoring: responding to emergencies (such as dropped tableware or spilled liquids).

[0140] Step 8: Large model / MoE adaptive iterative optimization and adjustment: If an abnormality occurs: the big model makes a decision: re-plan the subtask or overall strategy; the expert model fine-grainedly controls and corrects the action details; and quickly and adaptively handles task changes.

[0141] Step 9: Task completion and continuous learning: After completing the kitchen cleaning task: record the completed task process and any abnormal situations encountered in the knowledge base; the large model automatically summarizes the task experience, forms memories, and improves the performance of the next task.

[0142] This example illustrates the AI control system and robot capability enhancement: Foundation Model: Provides human-like general intelligence capabilities, including semantic understanding and task decomposition; general knowledge reasoning and decision-making; and real-time environmental understanding and adaptability.

[0143] Mixture of Experts (MoE): Multiple specialized domain models provide precise motion control and optimization: Safety Expert: ensures safe and reliable motion; Control Expert: ensures high-performance motion execution; Optimization Expert: optimizes motion planning paths in real time.

[0144] This embodiment, through the above processes and examples, demonstrates how a hybrid control system combining a large model and a MoE expert model can empower a general-purpose humanoid robot with high performance, adaptability, safety, reliability, and human-like intelligence, enabling intelligent management from specific task planning to action execution. This design ensures the robot's flexible and efficient operation in common scenarios, continuously iterating and upgrading its performance through task execution and real-time learning, truly moving toward the goal of an AGI robot. (See Table 2.)

[0145] Table 2 Advantages analysis of high-performance general humanoid robots Ability characteristics AI control implementation Advantages high performance Expert model fine motor control Accurate, fast and stable action execution Adaptability Large models adjust decisions in real time Quickly adapt to changes in environment and tasks Safety and reliability Security experts monitor and correct in real time Anti-collision and anti-fall to ensure safety Human-like intelligence General reasoning and learning for large models Possess general knowledge and learning ability

Claims

1. A high-performance general-purpose humanoid robot, characterized by: It includes a main body trunk equipped with a main control system, the main control system is provided with several standard communication interfaces and standard power supply interfaces, the main body trunk is connected to the upper limb platform through a first universal flange interface, the main body trunk is connected to the lower limb platform through a second universal flange interface, all actuators of the upper limb platform and the lower limb platform are connected to the standard power supply interface through a power supply line and to the standard communication interface through a communication line, and the main control system is provided with an overall coordinated control program based on deep reinforcement learning and model predictive control.

2. A high-performance universal humanoid robot according to claim 1, characterized in that: The main control system is electrically connected to an environment perception module for acquiring environment perception data around the robot.

3. A high-performance universal humanoid robot according to claim 2, characterized in that: The environment perception module includes a visual perception sensor, a force perception sensor, a tactile perception sensor, and an inertial measurement sensor.

4. A high-performance universal humanoid robot according to claim 2, characterized in that: The upper limb platform includes several seven-degree-of-freedom robotic arms of different load-bearing models, and the seven-degree-of-freedom robotic arms of different load-bearing models are all fixedly connected to the main body trunk through a first universal flange interface.

5. A high-performance universal humanoid robot according to claim 4, characterized in that: The robotic arm is connected to the execution end via a third universal flange interface, and the execution end includes dexterous hands or grippers of different models.

6. A high-performance universal humanoid robot according to claim 1, characterized in that: The lower limb platform includes a wheeled chassis or a footed chassis of different load-bearing models, and the wheeled chassis or the footed chassis of different load-bearing models are fixedly connected to the main body trunk through a second universal flange interface.

7. A high-performance universal humanoid robot according to claim 6, characterized in that: The wheeled chassis or foot-type chassis of different load-carrying models are equipped with 2-3 waist degrees of freedom.

8. A high-performance universal humanoid robot according to any one of claims 1 to 7, characterized in that: All actuators of the upper limb platform and lower limb platform include at least motors, encoders and drivers. The drivers all use EtherCAT communication protocol to communicate with standard communication interfaces through communication lines, and the motors all use unified power supply lines to connect to standard power supply interfaces for power supply.

9. A high-performance universal humanoid robot according to claim 2, 3, 4 or 5, characterized in that: The overall coordinated control program controls the upper limb platform including end position control and adaptive adjustment based on force feedback.

10. A high-performance universal humanoid robot according to claim 6 or 7, characterized in that: The overall coordination control program controls the lower limb platform by offline training of a gait generation model and online optimization and adjustment of step length and foothold position.

Images