A control method for an embodied robot that is independent of its body and position.

Abstract

This invention discloses a control method for an embodied robot that is independent of its body and position, comprising the following control steps: Step 1: Constructing a relativized motion space, modeling the robot's motion space as a relative incremental space of the end effector, including: end effector pose increment, outputting the pose increment of the current end effector relative to its own position at the previous moment; gripper control, outputting the opening / closing state or opening / closing degree increment; Step 2: Constructing a relativized body state space, using the robot's body center point as a reference point, and handling operation tasks and movement tasks respectively. This invention, by redesigning the motion space and body state space and adopting a relativized representation, enables the robot's strategy model to naturally adapt to different body structures and different initial positions, achieving zero-sample or few-sample transfer across bodies and positions.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically, to a control method that is independent of the robot body and its position. Background Technology

[0002] Current learning-based embodied robot control methods typically rely on large amounts of demonstration data specific to the robot body and its initial position for training. This data implicitly contains information about the robot's joint structure, link lengths, sensor mounting locations, and also its absolute world coordinates during task execution. This strong coupling leads to severe generalization problems in the trained models.

[0003] 1. Poor cross-entity generalization: A model trained on one robot cannot be directly transferred to another robot with a different structure. Even robots of the same type but slightly different in size require data to be collected again.

[0004] 2. Poor cross-position generalization: The model is sensitive to the robot's initial position. Once the initial position differs from the distribution in the training data, the control accuracy drops sharply.

[0005] 3. Low data utilization: Every time a new environment or a new robot is deployed, a large amount of data needs to be collected again, which is costly.

[0006] When performing tasks, humans do not rely on an absolute world coordinate system or precise values ​​of their joint angles. Humans perceive their own state and plan actions in a relative way: for example, we perceive the position of our hand relative to our body, and when performing an action, we focus on the relative relationship between the hand's current position and the target position, moving incrementally. This relative perception and incremental control mechanism allows humans to flexibly adapt to different body postures and starting positions without needing to relearn.

[0007] Therefore, there is an urgent need for a robot control method that draws on the principle of human relative perception, so as to decouple the robot's action representation and state representation from its body configuration and absolute position, thereby greatly improving the model's generalization ability. Summary of the Invention

[0008] To address the problems existing in the prior art, the present invention aims to provide a control method that is independent of the body and position of the embodied robot, which solves the technical problems of poor cross-body generalization, poor cross-position generalization, and low data utilization in the existing embodied robot control methods, and realizes zero-sample or small-sample transfer of robot strategy models across bodies and positions.

[0009] To solve the above problems, the present invention adopts the following technical solution;

[0010] A control method for an embodied robot that is independent of its body and position includes the following control steps:

[0011] Step 1: Construct a relativized motion space. Model the robot's motion space as a relative incremental space of the end effector, including: end effector pose increment, which outputs the pose increment of the current end effector relative to itself at the previous moment; gripper control, which outputs the opening / closing state or opening / closing degree increment.

[0012] Step 2: Construct a relativized ontological state space, using the robot's body center point as the reference point, to handle operational tasks and movement tasks separately. Specific task states are represented as follows:

[0013] Operation task status: Establish a coordinate system with reference points, and convert the absolute pose of each end point in the world coordinate system into a relative pose relative to its own coordinate system.

[0014] In this invention, the mobile task status is represented by an incremental representation of the body's center point.

[0015] Step 3: Based on the relativized action space and the relativized body state space, train the robot control strategy using either imitation learning or reinforcement learning methods. The input to the strategy network is the relativized body state and environmental observations, and the output is the end-effector incremental action and gripper control commands.

[0016] Step 4: Convert the incremental end effector output by the policy network into joint commands that the robot can execute. Add the increment to the current end effector pose to obtain the target end effector pose, and solve the target joint angles through inverse kinematics; the gripper control commands are directly mapped to gripper drive commands.

[0017] As a further description of the above technical solution:

[0018] In step two, the relative pose transformations of each end-effector in the task state are as follows:

[0019] ;

[0020] ;

[0021] in Let be the transformation matrix of the robot's body center point in the world coordinate system. Let be the position vector of the end effector wrist relative to the coordinate system of the body's center point. Let be the position vector of the end effector wrist in the world coordinate system. Let be the position vector of the end effector's ankle relative to the coordinate system of the body's center point. Let be the position vector of the end effector's ankle in the world coordinate system.

[0022] As a further description of the above technical solution:

[0023] In step two, the movement task state increment is represented by recording the pose change of the body's center point relative to its own coordinate system at the current moment:

[0024] ;

[0025] in This represents the pose change of the body's center point relative to its own coordinate system at the current moment. Let be the homogeneous transformation matrix of the body's center point in the world coordinate system at the current moment. It is the homogeneous transformation matrix of the body's center point in the world coordinate system at the previous moment.

[0026] As a further description of the above technical solution:

[0027] The architecture of the policy network in step three is as follows: input layer, relativized state vector; feature extraction layer, multilayer perceptron; output layer, terminal incremental action.

[0028] As a further description of the above technical solution:

[0029] In step three, the relativized body state includes collecting the robot's joint angles, angular velocities, and end effector forces; environmental observation includes visual images.

[0030] As a further description of the above technical solution:

[0031] The inverse kinematics solution formula in step four is as follows:

[0032] ;

[0033] ;

[0034] in This is the joint torque command vector. This is the joint velocity command vector. For Jacobian matrices, This is the pseudo-inverse of the Jacobian matrix. For the desired terminal velocity, It is an intermediate variable.

[0035] Compared with the prior art, the advantages of this invention are:

[0036] (1) In this scheme, by adopting the end-effector incremental motion space and the relativized ontology state space, the strategy model is completely decoupled from the robot's joint configuration, link length and other physical parameters. The trained model can be directly transferred to robots with different structures, realizing zero-sample cross-ontology generalization.

[0037] (2) In this scheme, the state representation excludes the absolute position information in the world coordinate system. The model is not sensitive to the robot's initial position and orientation, and can effectively perform tasks at any starting position, greatly improving deployment flexibility.

[0038] (3) In this scheme, since the coupling between the body and the position is decoupled, the same demonstration data can be used to train a general strategy applicable to multiple robots and multiple starting positions, avoiding repeated collection and greatly reducing data costs.

[0039] (4) This scheme draws on the mechanism of human relative perception and incremental control to make robot control more natural, robust, easy to understand and debug.

[0040] (5) This scheme designs state representations for operation and movement tasks respectively, and can be flexibly combined, making it suitable for complex robot platforms such as wheeled dual-arm robots. Attached Figure Description

[0041] Figure 1 This is a schematic diagram illustrating the control principle of an embodied robot that is independent of its position. Detailed Implementation

[0042] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0043] Please see Figure 1 The present invention provides the following embodiments:

[0044] A control method for an embodied robot that is independent of its body and position includes the following control steps:

[0045] Step 1: Construct a relativized motion space. Model the robot's motion space as a relative incremental space of the end effector, including: end effector pose increment, which outputs the pose increment of the current end effector relative to itself at the previous moment; gripper control, which outputs the opening and closing state or opening degree increment.

[0046] Step 2: Construct a relativized body state space, using the robot's body center point as the reference point, preferably the robot's hip joint as the body center point, and establish the hip joint coordinate system as its own coordinate system. The end effectors include the robot's wrist and ankle, which handle the manipulation task and the movement task respectively. The specific task state is represented as follows:

[0047] In the operation task state, a coordinate system is established with a reference point. The absolute pose of each end effector in the world coordinate system is converted into a relative pose relative to its own coordinate system. The relative pose transformation of each end effector is as follows:

[0048] ;

[0049] ;

[0050] in Let be the transformation matrix of the robot's body center point in the world coordinate system. Let be the position vector of the end effector wrist relative to the coordinate system of the body's center point. Let be the position vector of the end effector wrist in the world coordinate system. Let be the position vector of the end effector's ankle relative to the coordinate system of the body's center point. Let be the position vector of the end effector's ankle in the world coordinate system;

[0051] This representation eliminates the influence of the robot's overall displacement and orientation, allowing the model to focus on the spatial relationship of the end effector relative to the body center, independent of the robot's starting position. At the same time, this relative representation is also independent of the robot's specific joint lengths, because the end effector position is already a result of kinematic calculations, but it retains the spatial information required for the task.

[0052] In the context of mobile task status, this invention uses incremental representation of the body center point, that is, it records the pose change of the body center point relative to its own coordinate system at the current moment:

[0053] ;

[0054] in This represents the pose change of the body's center point relative to its own coordinate system at the current moment. Let be the homogeneous transformation matrix of the body's center point in the world coordinate system at the current moment. This is the homogeneous transformation matrix of the body's center point in the world coordinate system at the previous moment;

[0055] This incremental representation describes the robot's movement and turning, independent of absolute position and historical trajectory, enabling the model to learn general movement strategies unaffected by the starting position. For complex tasks that involve both manipulation and movement, both representations can be used simultaneously, depending on the task stage or network design, or the model can choose automatically.

[0056] For example, the hip joint increment can be used as global motion information, and the relative pose of the end effector can be used as local operation information, together forming the state space.

[0057] Step 3: Based on the relativized action space and the relativized body state space, train the robot control strategy using either imitation learning or reinforcement learning methods. The input to the strategy network is the relativized body state and environmental observations, and the output is the end-effector incremental action and gripper control commands.

[0058] The policy network architecture consists of: an input layer (relativized state vector), a feature extraction layer (multilayer perceptron), and an output layer (terminal incremental action).

[0059] Relative body state includes acquiring the robot's joint angles, angular velocities, and end effector forces; environmental observation includes visual images;

[0060] Since all inputs and outputs are decoupled from the ontology configuration and absolute position, the trained policy naturally possesses the following generalization capabilities:

[0061] Cross-body generalization allows the same strategy to be directly deployed to robots with different joint configurations. During deployment, the end-effector incremental commands are converted into joint control quantities based on the robot's kinematic model, and then solved through inverse kinematics.

[0062] Cross-location generalization means that the policy network is insensitive to the robot's starting position and can effectively perform tasks under different starting positions.

[0063] Step 4: Convert the incremental end effect output by the policy network into joint commands that the robot can execute. Add the increment to the current end effect pose to obtain the target end effect pose, and solve the target joint angles through inverse kinematics; the gripper control commands are directly mapped to gripper drive commands.

[0064] The formula for solving the inverse kinematics is as follows:

[0065] ;

[0066] ;

[0067] in This is the joint torque command vector. This is the joint velocity command vector. For Jacobian matrices, This is the pseudo-inverse of the Jacobian matrix. For the desired terminal velocity, It is an intermediate variable.

[0068] When actually deployed to a specific robot, the end-effector incremental motion output by the strategy needs to be converted into joint commands that the robot can execute. This process only requires the robot's kinematic model and does not require retraining the strategy.

[0069] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A control method for an embodied robot body and position-independent system, characterized in that, The following control steps are included: Step 1: Construct a relativized motion space. Model the robot's motion space as a relative incremental space of the end effector, including: end effector pose increment, which outputs the pose increment of the current end effector relative to itself at the previous moment; gripper control, which outputs the opening / closing state or opening / closing degree increment. Step 2: Construct a relativized ontological state space, using the robot's body center point as the reference point, to handle operational tasks and movement tasks separately. Specific task states are represented as follows: Operation task status: Establish a coordinate system with reference points, and convert the absolute pose of each end point in the world coordinate system into a relative pose relative to its own coordinate system. In this invention, the mobile task status is represented by an incremental representation of the body's center point. Step 3: Based on the relativized action space and the relativized body state space, train the robot control strategy using either imitation learning or reinforcement learning methods. The input to the strategy network is the relativized body state and environmental observations, and the output is the end-effector incremental action and gripper control commands. Step 4: Convert the incremental end effector output by the policy network into joint commands that the robot can execute. Add the increment to the current end effector pose to obtain the target end effector pose, and solve the target joint angles through inverse kinematics; the gripper control commands are directly mapped to gripper drive commands.

2. The control method for a bodyless robot and its position-independent body according to claim 1, characterized in that: In step two, the relative pose transformations of each end-effector in the task state are as follows: ； ； in Let be the transformation matrix of the robot's body center point in the world coordinate system. Let be the position vector of the end effector wrist relative to the coordinate system of the body's center point. Let be the position vector of the end effector wrist in the world coordinate system. Let be the position vector of the end effector's ankle relative to the coordinate system of the body's center point. Let be the position vector of the end effector's ankle in the world coordinate system.

3. The control method for a bodyless robot and its position-independent body according to claim 1, characterized in that: In step two, the movement task state increment is represented by recording the pose change of the body's center point relative to its own coordinate system at the current moment: ； in This represents the pose change of the body's center point relative to its own coordinate system at the current moment. Let be the homogeneous transformation matrix of the body's center point in the world coordinate system at the current moment. It is the homogeneous transformation matrix of the body's center point in the world coordinate system at the previous moment.

4. The control method for a bodyless robot and its position-independent body according to claim 1, characterized in that: The architecture of the policy network in step three is as follows: input layer, relativized state vector; feature extraction layer, multilayer perceptron; output layer, terminal incremental action.

5. The control method for a bodyless robot and its position-independent body according to claim 1, characterized in that: The relativistic body state in step three includes collecting the robot's joint angles, angular velocities, and end effector forces; Environmental observation includes visual images.

6. The control method for a bodyless robot and its position-independent body according to claim 1, characterized in that: The inverse kinematics solution formula in step four is as follows: ； ； in This is the joint torque command vector. This is the joint velocity command vector. For Jacobian matrices, This is the pseudo-inverse of the Jacobian matrix. For the desired terminal velocity, It is an intermediate variable.

Images