A mechanical arm dynamic path planning method and system based on analog learning

Abstract

The application provides a mechanical arm dynamic path planning method and system based on analog learning, and belongs to the technical field of robot path planning; the method comprises the following steps: for a mechanical arm grabbing task in a dynamic environment, obtaining a real-time environment state under a current grabbing task; based on an analog learning model, performing feature extraction on demonstration data when a human being performs a similar grabbing task, and generating a grabbing intention representation; according to the obtained real-time environment state and the grabbing intention representation, generating a plurality of candidate paths adapted to the current environment change; combining a reinforcement learning method and an obstacle avoidance detection method to determine an optimal path, and controlling the mechanical arm to perform a dynamic grabbing task according to the obtained optimal path; and the application can significantly improve the execution efficiency and the grabbing success rate of the mechanical arm in a dynamic environment through an analog learning and optimization mechanism.

Description

Technical Field

[0001] This invention belongs to the field of robot path planning technology, and particularly relates to a method and system for dynamic path planning of robotic arms based on simulation learning. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Robot path planning is a core component of autonomous robotic arm operations. In static environments, methods based on kinematics solving, sampling planning, or deep learning have achieved relatively stable path planning. However, as application scenarios expand to dynamic and unstructured environments, existing methods still have significant limitations when dealing with complex situations such as dynamic target movement and obstacle changes.

[0004] Specifically, existing technologies mainly have the following problems: (1) Existing methods are not responsive enough in dynamic environments and are difficult to update paths in real time. For example, existing path planning methods for guiding robotic arm movement mainly rely on predefined rules or offline calculations. When the environmental state changes, replanning is often required, which leads to a significant increase in planning delay and fails to meet the real-time requirements of dynamic tasks.

[0005] (2) In the prior art, the generated robotic arm motion path lacks integration with human operating experience, the trajectory is not natural enough, and the robustness under complex interference is poor. Although in recent years, methods have emerged to improve the anthropomorphism of the generated path by introducing imitation learning methods in robotic arm path planning, it is difficult to achieve continuous optimization and stable execution of the path in a dynamic environment by simply relying on imitation learning. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a dynamic path planning method and system for robotic arms based on simulation learning. Through simulation learning and optimization mechanisms, the execution efficiency and grasping success rate of robotic arms in dynamic environments can be significantly improved.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a dynamic path planning method for robotic arms based on simulation learning.

[0008] A dynamic path planning method for a robotic arm based on simulation learning includes: For robotic arm grasping tasks in dynamic environments, image data, depth information, and semantic descriptions of the target object and its surrounding environment are acquired to determine the real-time environmental state under the current grasping task. Based on the imitation learning model, feature extraction is performed on demonstration data of humans performing similar grasping tasks to generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic; Based on the obtained real-time environmental status and crawling intent, multiple candidate paths are generated to adapt to the current environmental changes; The comprehensive performance of each candidate path is evaluated using reinforcement learning, and obstacle avoidance detection is combined with obstacle avoidance detection to identify obstacle avoidance sections and determine the optimal path. Finally, the robotic arm is controlled to perform a dynamic grasping task according to the obtained optimal path.

[0009] Furthermore, the image data, depth information, and semantic description of the target object and its surrounding environment are acquired through a multimodal sensor, which includes an RGB-D camera and a force sensor.

[0010] Furthermore, the grasping intent representation includes the expected grasping posture of the target object, the robotic arm's movement priority, and the operational safety boundaries.

[0011] Furthermore, the generation of the candidate path includes: adjusting the number of candidate paths and the path search range according to the distribution density of obstacles and the movement speed of the target object in the real-time environment; increasing the number of candidate paths and narrowing the path search range when the obstacle density is high or the target object moves fast; and reducing the number of candidate paths and expanding the path search range when the obstacle density is low or the target object moves slowly, so as to improve the global optimality of the candidate path.

[0012] Furthermore, the reinforcement learning method evaluates the overall performance of each candidate path using a preset reward function; wherein the reward function uses capture success rate, energy consumption, trajectory smoothness, and obstacle avoidance constraints as evaluation indicators.

[0013] Furthermore, the evaluation indicators in the reward function adopt a scenario-adaptive allocation mechanism, namely: when obstacles in the environment change frequently, the weight of obstacle avoidance constraints is increased; when the target object is difficult to grasp, the weight of grasp success rate is increased; when the robotic arm needs to perform grasping tasks continuously for a long time, the weight of energy consumption during the movement process is increased.

[0014] Furthermore, the obstacle avoidance detection method includes: identifying road segments in the candidate path that may interfere with the obstacle by dynamically tracking the obstacle's position, and correcting them by local path replanning.

[0015] A second aspect of the present invention provides a dynamic path planning system for a robotic arm based on simulation learning.

[0016] A dynamic path planning system for a robotic arm based on simulation learning includes: The environment perception module is configured to: acquire image data, depth information and semantic description of the target object and its surrounding environment for robotic arm grasping tasks in dynamic environments, so as to determine the real-time environmental state under the current grasping task; The imitation learning module is configured to: extract features from demonstration data of humans performing similar grasping tasks based on the imitation learning model, and generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic; The path generation module is configured to generate multiple candidate paths that adapt to changes in the current environment based on the obtained real-time environmental status and crawling intent. The path optimization module is configured to: use reinforcement learning to evaluate the overall performance of each candidate path, and combine it with obstacle avoidance detection methods to identify obstacle avoidance segments and determine the optimal path; The execution module is configured to control the robotic arm to perform a dynamic grasping task according to the obtained optimal path. A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a simulation learning-based dynamic path planning method for a robotic arm as described in the first aspect of the present invention.

[0017] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a simulation learning-based dynamic path planning method for a robotic arm as described in the first aspect of the present invention.

[0018] The above one or more technical solutions have the following beneficial effects: (1) This invention extracts features from demonstration data of humans performing similar grasping tasks based on an imitation learning model to generate a grasping intent representation; then, based on the obtained real-time environmental state and grasping intent representation, it generates multiple candidate paths adapted to changes in the current environment. By rapidly generating multiple candidate paths and combining them with the real-time decision-making capabilities of reinforcement learning, the optimal solution is selected and locked. This mechanism enables the robotic arm to replan and switch paths in a very short time when it senses target movement or changes in obstacles, thereby greatly improving the real-time response speed and overall planning efficiency to dynamic environments, and effectively overcoming the delay problem caused by replanning in existing methods.

[0019] (2) This invention extracts features from demonstration data of humans performing similar grasping tasks using an imitation learning model and generates a grasping intent representation. Then, it combines reinforcement learning and obstacle avoidance detection methods for optimization. Imitation learning uses extracted human operational experience to guide path generation, ensuring that the initial path naturally possesses human-like smoothness and intuitive characteristics. Subsequently, reinforcement learning, based on this high quality, further optimizes and enhances robustness online to address uncertainties in dynamic environments. Under this collaborative mechanism, not only is the natural smoothness of the trajectory guaranteed, but the success rate of the robotic arm in complex dynamic environments and the overall robustness of the system are significantly improved, overcoming the limitations of single learning strategies in existing technologies.

[0020] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of the dynamic path planning method for the robotic arm in Embodiment 1 of the present invention.

[0023] Figure 2 This is a modular flowchart of the dynamic path planning system for the robotic arm in Embodiment 2 of the present invention. Detailed Implementation

[0024] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Overall approach: This invention provides a dynamic path planning method for robotic arms based on simulation learning. This method acquires image, depth, and semantic data of the target object and environment through multimodal sensors; extracts features from human demonstrations based on an imitation learning model to construct a grasping intent representation; generates candidate paths by combining environmental perception results; optimizes the candidate paths using reinforcement learning and collision detection methods to obtain a smooth and highly successful optimal path; and finally controls the robotic arm to perform dynamic grasping.

[0028] Example 1 This embodiment discloses a dynamic path planning method for robotic arms based on simulation learning.

[0029] like Figure 1 As shown, a dynamic path planning method for a robotic arm based on simulation learning includes: Step S1: For robotic arm grasping tasks in dynamic environments, acquire image data, depth information, and semantic descriptions of the target object and its surrounding environment to determine the real-time environmental state under the current grasping task. Step S2: Based on the imitation learning model, feature extraction is performed on the demonstration data of humans performing similar grasping tasks to generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic; Step S3: Based on the obtained real-time environmental status and crawling intent representation, generate multiple candidate paths that adapt to the current environmental changes; Step S4: Use reinforcement learning to evaluate the overall performance of each candidate path, and combine it with obstacle avoidance detection to identify obstacle avoidance sections and determine the optimal path; finally, control the robotic arm to perform a dynamic grasping task according to the obtained optimal path.

[0030] Based on the above process, this invention, through simulation learning and optimization mechanisms, can significantly improve the execution efficiency and grasping success rate of the robotic arm in dynamic environments. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0031] In step S1, for the robotic arm grasping task in a dynamic environment, image data, depth information and semantic description of the target object and its surrounding environment are acquired to determine the real-time environmental state under the current grasping task.

[0032] For robotic arm grasping tasks in dynamic environments, image data, depth information, and semantic descriptions of the target object and its surrounding environment are acquired; the dynamic environment includes target object movement, changes in obstacles in the scene, and external disturbances.

[0033] Image data, depth information, and semantic descriptions of the target object and its surrounding environment are acquired through multimodal sensors, including an RGB-D camera and a force sensor. Specifically, after installation and calibration, the RGB-D camera simultaneously acquires RGB images and depth maps of the target and its environment. After preprocessing and noise reduction, a semantic description containing object category and 3D coordinates is generated using Mask R-CNN. The six-axis force sensor is integrated into the wrist of the robotic arm. After calibration, it acquires force and torque data when the end effector approaches or contacts the object. After filtering, it corrects the deviation of the RGB-D depth information and combines force features to supplement semantic attributes such as object hardness and material. The two sensors are synchronized in time and space and verified with data to output fused perception data for subsequent dynamic path planning in the robotic arm's simulation learning.

[0034] In step S2, feature extraction is performed on demonstration data of humans performing similar grasping tasks based on the imitation learning model to generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic.

[0035] The imitation learning model is a behavior cloning model based on deep neural networks, specifically comprising four parts: an input layer, a feature extraction network (CNN+MLP), a feature fusion layer, and an action output layer. The input layer receives multimodal perception data, including RGB-D image and depth features, and force and torque features of force perception. In the feature extraction network, CNN processes image texture features, and MLP processes depth and force numerical features. The feature fusion layer integrates the two types of features. The action output layer generates the control quantities for the robotic arm. In practice, the input layer data is fed into the CNN to extract image edges and object contour features, and into the MLP to extract depth distance and force change features. These two types of features are weighted and concatenated by the fusion layer and then input into the output layer. The output layer maps the fused features to the joint angles of the robotic arm, achieving a direct mapping between perception and action.

[0036] Based on the imitation learning model, features are extracted from demonstration data of humans performing similar grasping tasks to generate a grasping intent representation that characterizes human operational habits and target grasping needs. Specifically, feature extraction from demonstration data of human grasping tasks based on the imitation learning model requires preprocessing the raw data first, using Kalman filtering or moving average to remove noise and unifying the time axis to achieve spatiotemporal alignment; then, multimodal features are extracted separately, joint motion data is processed using Transformer to obtain action temporal features, visual and scene features are extracted from the video using Visual Transformer, interaction features are extracted using MLP, and finally, multilayer perceptron is used to fuse multimodal features, training the model to fit human operational patterns using a behavior cloning approach. To generate a grasping intent representation, a mapping model with three independent output heads must first be constructed. The fused features are then input into the corresponding output heads: the posture output head predicts the Cartesian coordinates and posture angles of the end effector through regression, and corrects them using object visual features to match human habits; the motion priority output head relies on action temporal features, using a temporal sorting algorithm to determine the joint movement sequence and assign velocity weights; and the safety boundary output head combines scene and interaction force features, using a boundary regression model to predict spatial and force safety constraints. Finally, the intent parameters are simulated and verified in a virtual environment. If posture deviations or safety risks exist, the model parameters are adjusted in reverse for iterative optimization until the intent representation can stably guide the robotic arm to complete the grasping task, while consistently representing human operating habits and target grasping needs throughout the process. The grasping intent representation includes the expected grasping posture of the target object, the robotic arm's motion priority, and the operational safety boundaries.

[0037] In step S3, multiple candidate paths adapted to the current environmental changes are generated based on the obtained real-time environmental state and the crawling intent representation.

[0038] Based on the obtained real-time environmental state and grasping intent, multiple candidate paths adapted to the current environmental changes are generated. Specifically, after the model is trained, before using its weights for experimental testing, it is essential to ensure that the camera's fixed position and indexing remain consistent with the data acquisition process. This is because the camera's role in data acquisition is to obtain image information; the entire process relies on images for localization. Maintaining consistency between the camera position and indexing with the data acquisition process ensures that the data input into the imitation learning model network is valid, allowing the model to infer based on the trained weights and dynamically predict the robotic arm's trajectory. Once the required image data is input into the model network, combined with the current joint angle data of the robotic arm, the model can predict the K-step trajectory in real time. The hyperparameter K here is not a fixed value; it can be flexibly adjusted according to specific task requirements. For simple and fast sorting tasks, the K value can be set smaller so that the robotic arm can react quickly. For complex tasks requiring precise manipulation, the K value can be appropriately increased, allowing the model to have more predictions to plan a better motion path. Based on the current joint angle and camera sensor data, the next joint angle is predicted in real time, and the robot arm is controlled in real time to perform the pre-trained tasks through the robot arm control interface.

[0039] As an optional implementation, during path generation, the number of candidate paths and the path search range can be adjusted based on the obstacle distribution density and target object speed in the real-time environment. When the obstacle density is high or the target object moves quickly, the number of candidate paths is increased and the path search range is narrowed; when the obstacle density is low or the target object moves slowly, the number of candidate paths is decreased and the path search range is expanded to improve the global optimality of the candidate paths. Specifically, this can be achieved through the following method: First, a quantitative analysis of the real-time environment is performed. The obstacle distribution density can be calculated by the ratio of the number of obstacles detected in the current robotic arm workspace to their occupied volume; the target object speed is estimated in real-time by using visual tracking algorithms (such as optical flow or correlation filters) based on continuous frame image and depth data. Next, thresholds for the dynamic adjustment strategy are set. The system presets obstacle density thresholds for "high density" and "low density," and target movement speed thresholds for "fast" and "slow." When the system determines that the current environment has "high obstacle density" or "the target is moving quickly," the path generation module will activate a dense sampling mode. This means that, based on the grasping intent-guided feasible area, the number of candidate paths generated in parallel is increased (e.g., from the default 5 to 15). Simultaneously, the exploration step size and deflection angle of each path are limited, narrowing the path search range and concentrating it more on the forward space of the current optimal direction. This aims to quickly generate multiple short-term safe paths that avoid near-term collisions. Conversely, when the environment has "low obstacle density" and "the target is moving slowly or is stationary," the system reduces the number of candidate paths (e.g., to 3-5) and expands the path search range, allowing pathpoints to explore in a larger joint space or Cartesian space to find globally better paths with lower energy consumption or smoother trajectories. This adaptive mechanism, through real-time mapping of environmental states and generation strategies, ensures a balance between complexity and efficiency in the path planning process, thereby improving the global optimality of the candidate path set and its real-time scene adaptability.

[0040] In step S4, reinforcement learning is used to evaluate the overall performance of each candidate path, and obstacle avoidance detection is combined to identify obstacle avoidance sections and determine the optimal path. Finally, the robotic arm is controlled to perform a dynamic grasping task according to the obtained optimal path.

[0041] A reinforcement learning approach is employed to evaluate the overall performance of each candidate path, specifically through a pre-defined reward function. This reward function uses success rate, energy consumption, trajectory smoothness, and obstacle avoidance constraints as evaluation metrics. In the robotic arm grasping simulation of this embodiment, multiple reward functions are implemented to evaluate the robotic arm's grasping performance from different dimensions. The reward function acts as a bridge between the reinforcement learning algorithm and the actual control task, essentially transforming complex control objectives into quantifiable feedback signals. In robotic arm grasping tasks, success rate, grasping stability, and motion time are all important indicators for measuring grasping performance. By designing a successful grasping reward function, a positive reward is given when the robotic arm successfully grasps the target object and maintains stability; conversely, a negative reward is given if the grasp fails or the object falls. Furthermore, considering the dynamic characteristics of the grasping process, a motion smoothness reward function is introduced to evaluate the continuity and smoothness of the robotic arm's trajectory, avoiding unstable movements such as violent shaking or sudden stops and turns.

[0042] As an optional implementation, the evaluation metrics in the reward function adopt a scenario-adaptive allocation mechanism, namely: when obstacles in the environment change frequently, the weight of obstacle avoidance constraints is increased; when the target object is difficult to grasp, the weight of the grasp success rate is increased; when the robotic arm needs to perform grasping tasks continuously for a long time, the weight of energy consumption during the movement process is increased. The reward function can be formalized as follows: ; in, , , and These represent the dynamic weights corresponding to the success rate of capture, energy consumption, trajectory smoothness, and obstacle avoidance constraints, respectively. ; , , and These represent the scores for capture success rate, energy consumption, trajectory smoothness, and obstacle avoidance constraints after normalization.

[0043] Based on this, the scenario-adaptive allocation mechanism adopted by each evaluation metric in the reward function includes: 1) Obstacle avoidance constraint weights Adaptive mechanism: The system continuously monitors the positional changes of obstacles in the environment. It assesses the frequency or variance of obstacle position updates per unit time to evaluate the urgency of these dynamic changes. When this frequency exceeds a preset sensitivity threshold, the system automatically increases the obstacle avoidance constraint weights. (For example, increasing it from a baseline of 0.3 to 0.6) allows path optimization to prioritize safety.

[0044] 2) Weighting of crawling success rate Adaptive evaluation: The difficulty of grasping the target object is a comprehensive evaluation metric, based on factors including, but not limited to, the semantic type of the target object (e.g., fragile, irregular), the size and shape complexity calculated from point cloud data, and the success rate records of similar grasping tasks in the past. When the evaluator outputs "high" as the current grasping difficulty, the system increases the weight of the grasping success rate. This guides reinforcement learning models to tend to choose paths with more stable grasping postures and higher fault tolerance.

[0045] 3) Energy consumption weight Adaptive design: The system determines whether the robotic arm is in a "long-term continuous grasping task" working mode through task queues or host computer instructions. Simultaneously, it can also monitor the robotic arm's continuous working time or motor temperature rise. When a long-term task is determined, the system increases the weight of energy consumption during movement to reduce equipment wear and heat accumulation. This ensures that the optimized path minimizes unnecessary motion and torque output while guaranteeing success.

[0046] 4) Trajectory smoothness weight Adaptive: Given the obstacle avoidance constraint weight, capture success rate weight, and energy consumption weight, the trajectory smoothness weight can be obtained.

[0047] This adaptive allocation mechanism allows the core evaluation orientation of the reward function to be closely integrated with the characteristics of the dynamic environment and the task requirements, thereby ensuring that the path optimization results can reflect the most critical performance requirements in different scenarios.

[0048] Furthermore, the obstacle avoidance detection method includes: identifying road segments in candidate paths that may interfere with obstacles by dynamically tracking obstacle positions, and correcting them through local path replanning. In the specific implementation of the obstacle avoidance detection method, firstly, the obstacle positions are dynamically tracked using multimodal data acquired by the environmental perception module, and the movement state of the obstacles is captured through visual tracking. Simultaneously, auxiliary perception data is used to correct potential positioning deviations, and appropriate methods are employed to process data noise, ensuring real-time control of obstacle positions. Next, each candidate path is decomposed and analyzed to determine the safe distance between each road segment and the obstacle, marking road segments that may have interference risks, and initially screening candidate paths with excessively high risk ratios. Finally, local path replanning is performed on road segments with interference risks, reasonably limiting the scope of replanning, generating new local paths using appropriate algorithms, and smoothing the new paths to ensure smooth connection with the original paths. After replanning, the safety of the paths is verified again, and candidate paths that still cannot meet safety requirements are eliminated.

[0049] By combining reinforcement learning and obstacle avoidance detection methods, the optimal path is determined. In determining the optimal path using these methods, safety is the primary principle for initial screening of candidate paths, retaining those with no significant safety risks or whose risks are eliminated through local replanning. If the number of qualified paths is insufficient, a path generation module is triggered to generate additional candidate paths. Subsequently, the reward function in reinforcement learning is used to comprehensively evaluate the performance of the screened paths. The reward function considers key indicators such as grasping effectiveness, energy consumption, trajectory smoothness, and obstacle avoidance performance. Simultaneously, the evaluation weights of obstacle avoidance-related indicators are appropriately adjusted based on whether the path has undergone safety modifications, making the evaluation results more aligned with practical application needs. Finally, the candidate paths with the highest evaluation scores are further validated. The potential impact of subsequent environmental changes on the paths is predicted, and common environmental disturbances are simulated to assess the robustness of the paths. Combining the path's evaluation score and robustness performance, the final optimal path is determined, prioritizing paths with stronger stability in dynamic environments.

[0050] After determining the optimal path, the robotic arm is controlled to execute a dynamic grasping task according to the obtained optimal path. When controlling the robotic arm to execute the dynamic grasping task after determining the optimal path, the spatial information of the optimal path is first converted into control commands that the robotic arm joints can recognize and execute. These commands are then optimized based on the robotic arm's motion characteristics to avoid instability during movement. During execution, sensors continuously monitor the state of the target object, obstacles, and the robotic arm's own operation. If target deviation or obstacles intruding into the safe zone are detected, execution is paused immediately, and the path is locally adjusted. After adjustment, the task continues. When the robotic arm approaches the target object, it switches to a control mode more suitable for grasping operations to ensure smooth grasping action and maintain a stable grip after grasping. Simultaneously, corresponding handling strategies are formulated for different degrees of anomalies that may occur during execution. Minor anomalies can be resumed with brief adjustments, moderate anomalies require replanning the local path, and severe anomalies immediately stop execution and report fault information. After the task is completed, sensors confirm whether the target object has reached the expected position and state. If the task is confirmed to be completed, relevant parameters are reset to prepare for the next task. If the task is not completed, the cause of the problem is analyzed, relevant parameters are adjusted, and execution is retried.

[0051] Example 2 This embodiment discloses a dynamic path planning system for robotic arms based on simulation learning.

[0052] like Figure 2 As shown, a dynamic path planning system for a robotic arm based on simulation learning includes: The environment perception module is configured to: acquire image data, depth information and semantic description of the target object and its surrounding environment for robotic arm grasping tasks in dynamic environments, so as to determine the real-time environmental state under the current grasping task; The imitation learning module is configured to: extract features from demonstration data of humans performing similar grasping tasks based on the imitation learning model, and generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic; The path generation module is configured to generate multiple candidate paths that adapt to changes in the current environment based on the obtained real-time environmental status and crawling intent. The path optimization module is configured to: use reinforcement learning to evaluate the overall performance of each candidate path, and combine it with obstacle avoidance detection methods to identify obstacle avoidance segments and determine the optimal path; The execution module is configured to control the robotic arm to perform a dynamic grasping task according to the obtained optimal path. Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0053] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a simulation learning-based dynamic path planning method for a robotic arm as described in Embodiment 1 of this disclosure.

[0054] Example 4 The purpose of this embodiment is to provide an electronic device.

[0055] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in a simulation learning-based dynamic path planning method for a robotic arm as described in Embodiment 1 of this disclosure.

[0056] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0057] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0058] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A dynamic path planning method for a robotic arm based on simulation learning, characterized in that, include: For robotic arm grasping tasks in dynamic environments, image data, depth information, and semantic descriptions of the target object and its surrounding environment are acquired to determine the real-time environmental state under the current grasping task. Based on the imitation learning model, feature extraction is performed on demonstration data of humans performing similar grasping tasks to generate a grasping intent representation that characterizes human operating habits and target grasping needs. The demonstration data includes hand movement trajectory, grasping posture adjustment, and environmental interaction logic. The generation of the grasping intent representation is achieved by constructing a mapping model containing three independent output heads: a posture output head that predicts the end Cartesian coordinates and posture angles through regression; a motion priority output head that determines the joint movement sequence and assigns velocity weights using a temporal sorting algorithm; and a safety boundary output head that predicts the safety constraints of space and force through a boundary regression model. Based on the obtained real-time environmental status and crawling intent, multiple candidate paths are generated to adapt to the current environmental changes; A reinforcement learning method is used to evaluate the overall performance of each candidate path, while an obstacle avoidance detection method is combined to identify obstacle avoidance sections and determine the optimal path. The reinforcement learning method evaluates the overall performance of each candidate path using a pre-defined reward function. This reward function uses grasping success rate, energy consumption, trajectory smoothness, and obstacle avoidance constraints as evaluation metrics, and each evaluation metric employs a scene-adaptive allocation mechanism. Finally, the robotic arm is controlled to execute a dynamic grasping task according to the obtained optimal path. The reward function is expressed as follows: ； In the formula, , , and These represent the dynamic weights corresponding to the success rate of capture, energy consumption, trajectory smoothness, and obstacle avoidance constraints, respectively. ; , , and These represent the scores for capture success rate, energy consumption, trajectory smoothness, and obstacle avoidance constraints after normalization.

2. The method for dynamic path planning of a robotic arm based on simulation learning as described in claim 1, characterized in that, The image data, depth information, and semantic description of the target object and its surrounding environment are acquired through a multimodal sensor, which includes an RGB-D camera and a force sensor.

3. The dynamic path planning method for a robotic arm based on simulation learning as described in claim 1, characterized in that, The grasping intent representation includes the expected grasping posture of the target object, the robotic arm's movement priority, and the operational safety boundaries.

4. The dynamic path planning method for a robotic arm based on simulation learning as described in claim 1, characterized in that, The generation of the candidate path includes: adjusting the number of candidate paths and the path search range according to the distribution density of obstacles and the movement speed of the target object in the real-time environment; increasing the number of candidate paths and narrowing the path search range when the obstacle density is high or the target object moves fast; and reducing the number of candidate paths and expanding the path search range when the obstacle density is low or the target object moves slowly, so as to improve the global optimality of the candidate path.

5. The dynamic path planning method for a robotic arm based on simulation learning as described in claim 1, characterized in that, The evaluation indicators in the reward function adopt a scenario-adaptive allocation mechanism, namely: when obstacles in the environment change frequently, the weight of obstacle avoidance constraints is increased; when the target object is difficult to grasp, the weight of grasp success rate is increased; when the robotic arm needs to perform grasping tasks continuously for a long time, the weight of energy consumption during the movement process is increased.

6. The dynamic path planning method for a robotic arm based on simulation learning as described in claim 1, characterized in that, The obstacle avoidance detection method includes: identifying road segments in the candidate path that may interfere with the obstacle by dynamically tracking the obstacle's position, and correcting them by local path replanning.

7. A robotic arm dynamic path planning system based on simulation learning, employing the robotic arm dynamic path planning method as described in any one of claims 1-6, characterized in that, include: The environment perception module is configured to: acquire image data, depth information and semantic description of the target object and its surrounding environment for robotic arm grasping tasks in dynamic environments, so as to determine the real-time environmental state under the current grasping task; The imitation learning module is configured to: extract features from demonstration data of humans performing similar grasping tasks based on the imitation learning model, and generate a grasping intent representation that characterizes human operating habits and target grasping needs; wherein, the demonstration data includes hand movement trajectory, grasping posture adjustment and environmental interaction logic; The path generation module is configured to generate multiple candidate paths that adapt to changes in the current environment based on the obtained real-time environmental status and crawling intent. The path optimization module is configured to: use reinforcement learning to evaluate the overall performance of each candidate path, and combine it with obstacle avoidance detection methods to identify obstacle avoidance segments and determine the optimal path; The execution module is configured to control the robotic arm to perform a dynamic grasping task according to the obtained optimal path.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the simulation learning-based dynamic path planning method for a robotic arm as described in any one of claims 1-6.

9. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the simulation learning-based dynamic path planning method for a robotic arm as described in any one of claims 1-6.

Images