Vision-based stateless robot motion control method and system

Abstract

The invention discloses a stateless robot motion control method and system based on vision, and belongs to the field of robot intelligent control, and the method comprises the steps: collecting multi-view visual observation data through a top main camera and double-wide-angle wrist cameras installed above and below an end effector; inputting the visual observation data into a pre-trained strategy network, wherein the strategy network outputs a relative pose increment of an end effector; and the current pose of the end effector is updated according to the relative pose increment, and the robot is controlled to execute corresponding actions. The strategy network is trained based on a stateless imitation learning framework, only visual observation data and corresponding expert action labels are used in the training process, and state information is sensed without depending on a robot body; according to the method, pure-vision-driven end-to-end control is realized, and the method has strong space generalization ability, high data efficiency and excellent cross-platform adaptability, and is suitable for complex operation tasks such as industrial manufacturing and service robots.

Description

Technical Field

[0001] This invention relates to the field of robot motion control technology, and in particular to a vision-based stateless robot motion control method and system. Background Technology

[0002] Visual motion control technology has become an important direction in the development of robot intelligence in recent years. It establishes a direct mapping from visual observation to robot actions through imitation learning or reinforcement learning, enabling autonomous operation without human intervention. Traditional visual motion control methods typically employ a joint strategy structure of "visual input + proprioceptive state input." While providing environmental information from visual sensors, it also incorporates internal state information such as joint angles and end effector poses to improve the accuracy and stability of the movements.

[0003] For example, existing research has integrated data such as end effector position, joint angle state, and external force feedback into a vision module to achieve tasks such as complex assembly, folding, and object handling. This type of method improves control accuracy in known environments by combining visual observation with proprioceptive state information. However, such methods still have the following significant drawbacks in practical applications:

[0004] Limited generalization ability: Because policy networks tend to memorize the mapping relationship between specific states and actions during training, when the spatial position or height of the target object in the task scene changes, the model has difficulty adapting to the new layout, resulting in a sharp drop in performance.

[0005] Strong data dependence: In order to avoid overfitting, a large amount of demonstration data under different spatial layouts needs to be collected to cover the diversity of state space. In real robot systems, data collection is costly and inefficient.

[0006] Cross-platform migration is difficult: The ontological state spaces (such as joint configuration and pose representation) of different robots differ, making it difficult to directly transfer trained strategies to other robot platforms, thus limiting their versatility and scalability.

[0007] Although existing research has attempted to alleviate the aforementioned problems through methods such as data augmentation and state normalization, it has not fundamentally eliminated the dependence on the robot's internal state, making it difficult to achieve a balance between high spatial generalization, low data dependency, and cross-platform adaptability while maintaining accuracy. Therefore, how to design a robot motion control method with strong generalization ability, high data efficiency, and good transferability, relying solely on visual perception, has become a pressing technical problem to be solved in this field. Summary of the Invention

[0008] The problem solved by this invention is to provide a vision-based stateless robot motion control method and system, which realizes end-to-end control driven by pure vision, has strong spatial generalization ability, high data efficiency and excellent cross-platform adaptability, and is suitable for complex operation tasks such as industrial manufacturing and service robots.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a vision-based stateless robot motion control method, comprising the following steps: S1 acquiring visual observation data at the current moment, wherein the visual observation data is collected by a top main camera and at least one wrist camera; S2 inputting the visual observation data into a pre-trained policy network, wherein the policy network outputs the relative pose increment of the end effector; S3 updating the current pose of the end effector according to the relative pose increment to control the robot to perform corresponding actions; wherein the policy network is trained based on a stateless imitation learning framework, and during the training process only visual observation data and its corresponding expert action labels are used, and robot body perception state information is not used.

[0010] Preferably, the relative pose increment is expressed as:

[0011] ,

[0012] in, For translation increment, As the attitude increment, the updated pose of the end effector is obtained through pose composition operation.

[0013] Preferably, the wrist camera includes two wide-angle cameras, respectively mounted above and below the end effector, with the wrist camera facing the same direction as the end effector.

[0014] Preferably, the field of view of both wide-angle cameras is 120°×120°.

[0015] Preferably, the policy network is trained by minimizing the negative log-likelihood loss, and the loss function is: ,

[0016] in, For visual observation data, Tag for expert actions, This represents the action distribution output by the policy network.

[0017] A vision-based stateless robot motion control system includes: an A1 vision acquisition module for acquiring visual observation data at the current moment, including a top main camera and at least one wrist camera; an A2 policy network module for receiving the visual observation data and outputting the relative pose increment of the end effector; and an A3 control execution module for updating the current pose of the end effector according to the relative pose increment and controlling the robot to perform corresponding actions. The policy network module is trained based on a stateless imitation learning framework, using only visual observation data and its corresponding expert action labels during training, and does not use the robot's own perceived state information.

[0018] Preferably, the field of view of both wide-angle cameras is 120°×120°.

[0019] Preferably, when the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

[0020] A robot comprising the system as described in any one of claims 6 to 8.

[0021] The beneficial effects of this invention are:

[0022] Strong spatial generalization ability: By adopting a relative end-effector action space, the policy network predicts relative pose increments based solely on visual input, rather than absolute pose, enabling consistent action outputs from the same visual observation at different spatial heights and horizontal positions. In the example task of putting a pen into a pen case, the method improves the generalization success rate to 0.98 in the vertical direction and 0.58 in the horizontal direction, significantly outperforming existing methods that rely on state input.

[0023] High data efficiency: By eliminating the dependence on the robot's internal state, the strategy no longer needs to remember diverse state trajectory mappings, thus reducing the reliance on large-scale demonstration data. Experiments show that, taking folding clothes as an example, a success rate of over 80% can be achieved using only 50 demonstration data points, significantly reducing data collection costs and time overhead.

[0024] Excellent cross-platform adaptability: Since control is entirely based on visual input, it avoids domain offset issues caused by differences in the robot's state space. When migrating between different robot shapes or joint configurations, only minor adjustments to the image input are needed to quickly adapt the strategy, significantly improving the method's versatility and deployment flexibility.

[0025] Comprehensive and stable visual input: By adopting a collaborative observation architecture of a top main camera and upper and lower dual wide-angle wrist cameras, full-task visual coverage from global to local and from top to bottom is achieved, effectively solving the problems of limited field of view and severe occlusion of traditional wrist cameras, and enhancing the robustness and information integrity of the visual system.

[0026] End-to-end stateless control: The entire control process does not rely on any internal state information of the robot, realizing pure vision-driven end-to-end imitation learning and control, simplifying system architecture, improving deployment convenience, and applicable to various robot platforms and complex task scenarios. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] Specific implementation examples are given below.

[0029] A vision-based stateless robot motion control method includes the following steps: S1 acquiring visual observation data at the current moment, the visual observation data being collected by a top main camera and at least one wrist camera; S2 inputting the visual observation data into a pre-trained policy network, the policy network outputting the relative pose increment of the end effector; S3 updating the current pose of the end effector according to the relative pose increment to control the robot to perform corresponding actions; wherein, the policy network is trained based on a stateless imitation learning framework, and during the training process, only visual observation data and its corresponding expert action labels are used, and robot body perception state information is not used;

[0030] The relative pose increment is expressed as:

[0031] ,

[0032] in, For translation increment, As the attitude increment, the updated pose of the end effector is obtained through pose composition operation;

[0033] The wrist camera includes two wide-angle cameras, which are mounted above and below the end effector, respectively, with the wrist cameras facing the same direction as the end effector;

[0034] Both wide-angle cameras have a field of view of 120° × 120°;

[0035] The policy network is trained by minimizing the negative log-likelihood loss, with the loss function being: ,

[0036] in, For visual observation data, Tag for expert actions, The action distribution output by the policy network;

[0037] A vision-based stateless robot motion control system includes: an A1 vision acquisition module for acquiring visual observation data at the current moment, including a top main camera and at least one wrist camera; an A2 policy network module for receiving visual observation data and outputting the relative pose increment of the end effector; and an A3 control execution module for updating the current pose of the end effector according to the relative pose increment and controlling the robot to perform corresponding actions. The policy network module is trained based on a stateless imitation learning framework, using only visual observation data and its corresponding expert action labels during training, without using the robot's own perceived state information.

[0038] The wrist camera includes two wide-angle cameras, which are mounted above and below the end effector, respectively, with the wrist cameras facing the same direction as the end effector;

[0039] Both wide-angle cameras have a field of view of 120° × 120°;

[0040] When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5;

[0041] Strong spatial generalization ability: By adopting a relative end effector action space, the policy network predicts relative pose increments based solely on visual input, rather than absolute pose, enabling consistent action outputs from the same visual observation at different spatial heights and horizontal positions. In the example task of putting a pen into a pen case, the method improves the generalization success rate to 0.98 in the vertical direction and 0.58 in the horizontal direction, which is significantly better than existing methods that rely on state input.

[0042] High data efficiency: By eliminating the dependence on the robot's internal state, the strategy no longer needs to remember diverse state trajectory mappings, thereby reducing the reliance on large-scale demonstration data. Experiments show that, taking folding clothes as an example, in this task, a success rate of over 80% can be achieved using only 50 demonstration data points, significantly reducing data collection costs and time overhead.

[0043] Excellent cross-platform adaptability: Since control is based entirely on visual input, it avoids the domain offset problem caused by differences in the robot's state space; when migrating between different robot shapes or joint configurations, only slight adjustments to the image input are needed to quickly adapt the strategy, greatly improving the versatility and deployment flexibility of the method.

[0044] Comprehensive and stable visual input: By adopting a collaborative observation architecture of a top main camera and upper and lower dual wide-angle wrist cameras, full-task visual coverage from global to local and from top to bottom is achieved, effectively solving the problems of limited field of view and severe occlusion of traditional wrist cameras, and enhancing the robustness and information integrity of the visual system.

[0045] End-to-end stateless control: The entire control process does not rely on any internal state information of the robot, realizing pure vision-driven end-to-end imitation learning and control, simplifying system architecture, improving deployment convenience, and applicable to various robot platforms and complex task scenarios.

[0046] The above description is only a preferred embodiment of the present invention, but 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 inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A vision-based stateless robot motion control method, characterized in that, Includes the following steps: S1 Acquires visual observation data at the current moment, which is collected by the top main camera and at least one wrist camera; S2 Inputs the visual observation data into a pre-trained policy network, which outputs the relative pose increment of the end effector; S3 Updates the current pose of the end effector according to the relative pose increment to control the robot to perform corresponding actions; wherein, the policy network is trained based on a stateless imitation learning framework, and only uses visual observation data and its corresponding expert action labels during the training process, without using robot body perception state information.

2. The vision-based stateless robot motion control method according to claim 1, characterized in that, The relative pose increment is expressed as: ， in, For translation increment, As the attitude increment, the updated pose of the end effector is obtained through pose composition operation.

3. The vision-based stateless robot motion control method according to claim 1, characterized in that, The wrist camera includes two wide-angle cameras, which are mounted above and below the end effector, respectively, and the wrist camera is oriented in the same direction as the end effector.

4. The vision-based stateless robot motion control method according to claim 1, characterized in that, Both wide-angle cameras have a field of view of 120° × 120°.

5. The vision-based stateless robot motion control method according to claim 1, characterized in that, The policy network is trained by minimizing the negative log-likelihood loss, the loss function of which is: , in, For visual observation data, For expert action tags, This represents the action distribution output by the policy network.

6. A vision-based stateless robot motion control system, characterized in that, include: The A1 visual acquisition module is used to acquire visual observation data at the current moment, including the top main camera and at least one wrist camera; The A2 strategy network module is used to receive the visual observation data and output the relative pose increment of the end effector. The A3 control execution module is used to update the current pose of the end effector according to the relative pose increment and control the robot to perform corresponding actions. The policy network module is trained based on a stateless imitation learning framework. During the training process, only visual observation data and its corresponding expert action labels are used, and the robot's own perception state information is not used.

7. A vision-based stateless robot motion control system according to claim 1, characterized in that, Both wide-angle cameras have a field of view of 120° × 120°.

8. A vision-based stateless robot motion control system according to claim 1, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

9. A robot, characterized in that, The system includes any one of claims 6 to 7.

Images