Mechanical arm path planning method based on multi-source trajectory fusion and learning strategy

Abstract

The invention relates to the technical field of robot path planning, in particular to a mechanical arm path planning method based on multi-source trajectory fusion and a learning strategy. According to the method, firstly, multiple traditional path planning algorithms are utilized to generate multi-source candidate trajectories, and uniform representation of the trajectories is realized through path progress alignment and spatial resampling; and then a multi-observation trajectory fusion model is constructed based on uncertainty modeling, and recursive fusion is performed on the multi-source trajectory by using a Kalman filtering method to obtain a stable and reliable fusion trajectory. And further taking the fusion trajectory as expert demonstration data to train a path generation strategy, and optimizing the strategy in combination with reinforcement learning to realize autonomous generation and optimization of the mechanical arm path in a complex environment. The method can make full use of the information advantages of different planning algorithms, improves the stability and safety of path planning, improves the training efficiency and environment adaptability of a learning strategy, and is suitable for robot path planning and autonomous control tasks in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of robot path planning technology, specifically, it relates to a robotic arm path planning method based on multi-source trajectory fusion and learning strategies. Background Technology

[0002] With the rapid development of intelligent manufacturing and robotics, industrial robotic arms are widely used in fields such as electronic assembly, automated production, and logistics handling. In actual task execution, robotic arms need to generate safe, feasible, and efficient motion paths in complex environments with obstacles. Therefore, path planning technology has become a key issue in robotic systems, and its performance directly affects the working efficiency and safety of the robotic system.

[0003] Currently, path planning methods for robotic arms mainly include traditional algorithm-based methods and learning-based methods. Traditional path planning algorithms, such as graph search methods like A* and Dijkstra's algorithm, can search for feasible paths in discrete space and have the advantage of high stability. Sampling methods, such as RRT and PRM, explore the feasible space through random sampling and have good adaptability in high-dimensional environments. However, single planning algorithms usually have fixed search strategies, making it difficult to simultaneously consider multiple indicators such as path length, smoothness, and safety in complex environments, thus limiting the stability and adaptability of the planning results.

[0004] In recent years, learning methods such as imitation learning and reinforcement learning have been increasingly applied to robot path planning tasks. These methods establish a mapping relationship between environmental states and path decisions by learning from expert demonstration data, thereby improving the robot's autonomous planning capabilities in complex environments. However, these methods typically rely on high-quality expert demonstration data; when the quality of the demonstration data is insufficient or its distribution is inadequate, the performance of the learning strategy is prone to degradation.

[0005] Therefore, how to effectively utilize multi-source path information generated by various path planning algorithms and construct high-quality demonstration data while ensuring path stability, thereby further improving the robot's path planning capability, has become an important issue in current robot path planning research.

[0006] In view of this, the present invention is proposed. Summary of the Invention

[0007] To solve the above-mentioned technical problems, the basic concept of the technical solution adopted by the present invention is as follows:

[0008] To address the shortcomings of existing robotic arm path planning methods in complex environments, such as insufficient path stability, limited adaptability of single planning algorithms, and reliance on high-quality demonstration data for learning methods, this invention proposes a robotic arm path planning method based on multi-source trajectory fusion and a learning strategy. This method improves the stability, robustness, and environmental adaptability of path planning by fusing trajectory information generated by multiple path planning algorithms and incorporating a learning strategy for optimization.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention proposes a path planning method for robotic arms based on multi-source trajectory fusion and learning strategies, characterized in that the method includes the following steps.

[0011] 1. Multi-source path generation

[0012] Given a starting point and a target point, multiple candidate paths are generated using various path planning algorithms. These algorithms include graph search-based methods and sampling-based methods. The paths generated by different algorithms can be represented as follows:

[0013]

[0014] in Indicates the first The path generated by the planning algorithm This indicates the location of a path point in three-dimensional space.

[0015] 2. Path alignment and resampling

[0016] Since different path planning algorithms generate different numbers of path points and sampling densities, a unified representation of the path is needed. First, the path progress is normalized using the path arc length:

[0017]

[0018] in This represents the normalized progress of the path execution. Subsequently, spatial resampling is performed within a unified progress interval to uniformly represent the path as a fixed-length path sequence.

[0019]

[0020] This ensures that the paths generated by different planning algorithms maintain structural consistency.

[0021] 3. Multi-source trajectory fusion

[0022] After completing path alignment and resampling, the paths generated by different path planning algorithms are considered as multi-source observations of the real path state. Let the fused path be at the progress position... The actual state is , No. The observations of the path at this location are as follows:

[0023]

[0024] in To observe the noise, the multi-source paths are fused based on the Kalman filter method. The prediction and update process is as follows:

[0025] predict:

[0026]

[0027] renew:

[0028]

[0029] in For Kalman gain, This represents the observation noise covariance corresponding to different path sources. The fusion path is obtained through recursive updating.

[0030]

[0031] 4. Integration of trajectory-driven imitation learning

[0032] The fused paths are used as expert demonstration data to construct the training dataset. A policy network is generated by imitating these paths, enabling the model to learn the mapping relationship from states to path target points. Its loss function is defined as:

[0033]

[0034] in This is the robot's current state. These are the target path points in the expert trajectory. For policy networks.

[0035] 5. Optimize reinforcement learning strategies

[0036] Based on the initialization strategy of imitation learning, the path generation strategy is further optimized by interacting with the environment through reinforcement learning methods, thereby improving the robot's path planning ability and environmental adaptability in complex environments.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] By generating multiple candidate paths through multi-source path planning, the complementary advantages of different planning algorithms in search mechanisms can be fully utilized to improve the stability of path planning.

[0039] By aligning paths and resampling, different paths are kept structurally consistent, providing a unified data foundation for multi-source fusion.

[0040] A trajectory fusion method based on Kalman filtering is adopted to achieve adaptive fusion of multi-source path information based on uncertainty modeling, thereby improving the smoothness and reliability of the fused path.

[0041] By combining trajectory fusion with imitation learning and reinforcement learning methods, we can improve the efficiency of policy training and enhance the robot's autonomous planning ability in complex environments.

[0042] Therefore, the path planning method proposed in this invention can significantly improve the success rate, safety and stability of robotic arm path planning, and has good engineering application value.

[0043] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0044] In the attached diagram:

[0045] Figure 1 This is a flowchart of the implementation method of the present invention.

[0046] Figure 2 This is a performance comparison chart of the present invention with existing path planning methods in terms of path success rate.

[0047] Figure 3 This is a performance comparison chart of the present invention with existing path planning methods in terms of path length.

[0048] Figure 4 This is a performance comparison chart of the present invention with existing path planning methods in terms of path smoothness.

[0049] Figure 5 This paper compares the performance of the present invention with existing path planning methods in terms of path safety margin. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention.

[0051] To verify the effectiveness and robustness of the proposed multi-source path fusion method under different information source conditions, this paper constructs a multi-source ensemble containing four graph search, seven sampling, one potential field, and one optimization expert path planning methods to generate diverse and complementary trajectory data. This expert ensemble covers three typical path planning paradigms: graph search, sampling, and potential field, and can comprehensively reflect the behavioral differences of different planning strategies in three-dimensional space, providing representative multi-source inputs for subsequent fusion algorithms.

[0052] Table 1 Key parameter configurations for existing path planning methods

[0053] Expert Name category Key parameter configuration Select the destination baseline path A* Image search Heuristic function Stable, high quality Dijkstra Image search No heuristic function Graph search class stable baseline Weighted A* Image search Heuristic weight w=1.2 Improve search speed and introduce suboptimal paths Theta* Image search Enabling arbitrary angle search Generates smoother paths and provides stylistic variation. RRT-s sampling step size = 0.02, goal bias = 0.05 Smaller steps allow for more precise path design but slower exploration. RRT-l sampling step size = 0.08, goal bias = 0.05 Large strides, radical exploration but with a relatively crude approach RRT-Connect sampling Bidirectional expansion enabled Improve search efficiency and introduce structural differences RRT* sampling Reconnection mechanism enabled Asymptotically optimal path, higher quality PRM-200 sampling Sampling points N = 200, k = 10 Construct a sparse road network with coarse paths PRM-500 sampling Sampling points N = 500, k = 15 Building a dense road network leads to better routes. PRM* sampling Asymptotic optimal connection strategy Provide high-quality structured paths APF Power Field The attraction coefficient η_att = 1.0, and the repulsion coefficient η_rep = 2.5. As a weak expert, verifying robustness CHOMP optimization Optimize α = 0.02, λ_smooth = 1.0, λ_obs = 50 Continuous optimization generates smooth paths, supplementing the expert style of optimization.

[0054] The multi-source expert system constructed in this paper has the following significant advantages. First, the expert ensemble exhibits ample diversity, encompassing not only the structural differences between different planning paradigms (graph search, sampling methods, and potential field methods) but also the behavioral differences within the same class of algorithms due to varying parameter configurations. This provides more representative multi-source information input. This diverse design helps to comprehensively reflect the path planning needs in complex environments and lays the foundation for validating the effectiveness of subsequent fusion strategies. The system was trained for 300 epochs.

[0055] 2. Algorithm Evaluation Metrics

[0056] To comprehensively evaluate the overall performance of different path planning algorithms and their fusion strategies in terms of path quality and algorithm performance, this paper constructs a systematic quantitative evaluation index system from multiple dimensions, including feasibility, geometric characteristics, security, and execution controllability. This system not only focuses on whether the algorithm can successfully generate feasible paths, but also delves into key factors such as the geometric efficiency, smoothness, and safety margin of the generated paths, thus providing multi-faceted experimental support for the effectiveness and practicality of the methods.

[0057] First, the success rate (SR) measures an algorithm's ability to generate collision-free feasible paths given a set of tasks. It is defined as the ratio of successfully completed tasks to the total number of tasks.

[0058]

[0059] in This represents the number of tasks that successfully planned collision-free paths. This represents the total number of test tasks. This metric directly reflects the robustness and reliability of the algorithm in complex environments.

[0060] Secondly, to evaluate the geometric efficiency of a path, this paper introduces path length. As one of the core indicators, the planned path consists of T discrete points, and its length is defined as the sum of the Euclidean distances between adjacent path points:

[0061]

[0062] in This represents the spatial coordinates of the path at time t. A shorter path generally means higher motion efficiency.

[0063] Furthermore, smoothness The curvature variation of a path, used to characterize its changes, significantly impacts the smoothness and energy consumption of robot motion. This paper employs a smoothness measure based on second-order differences.

[0064]

[0065] The smaller this indicator is, the more gradual the change in local curvature of the path, the smoother the trajectory, and the more beneficial it is for subsequent trajectory tracking and control.

[0066] In terms of safety, this paper defines the minimum obstacle distance (Clearance). To measure the safety margin between the path and obstacles in the environment:

[0067]

[0068] in Representing path points To the obstacle collection The shortest Euclidean distance. Larger. A higher value implies a greater safety margin for obstacle avoidance, which helps improve the system's fault tolerance in dynamic or uncertain environments.

[0069] In summary, the above four indicators together form the basis for this paper's systematic evaluation of path planning and fusion methods. They cover both algorithmic performance and engineering requirements in actual deployment, providing a comprehensive and objective quantitative basis for subsequent experimental analysis.

[0070] 3. Comparison of hand-eye calibration algorithm results

[0071] Based on the method proposed in this invention, a total of 100 experiments were conducted and compared with existing path planning methods. The comparison results are as follows:

[0072] Table 2 shows the comparison results between the proposed algorithm and existing methods.

[0073] Method SR PL(m) PS (rad / m) (m) A* 0.82 1.41 0.72 0.061 Dijkstra 0.84 1.48 0.78 0.063 Weighted A* 0.83 1.37 0.91 0.059 Theta* 0.86 1.33 0.66 0.064 RRT-s 0.78 1.66 0.95 0.057 RRT-l 0.82 1.58 0.84 0.060 RRT-Connect 0.88 1.50 0.71 0.062 RRT* 0.86 1.36 0.63 0.066 PRM-s 0.80 1.61 0.92 0.058 PRM-l 0.83 1.55 0.83 0.060 PRM* 0.87 1.45 0.70 0.065 APF 0.77 1.46 0.89 0.052 CHOMP 0.90 1.34 0.54 0.071 KF-Fusion+Learning 0.91 1.32 0.50 0.075

[0074] .

Claims

1. A path planning method for a robotic arm based on multi-source trajectory fusion and learning strategies, characterized in that, The method includes: Acquire the robot's initial pose, target pose, and environmental obstacle information in the robot's motion environment; Based on the environmental information, multiple candidate paths are generated using various path planning algorithms; The paths generated by different path planning algorithms are uniformly represented to eliminate the differences between the number of path points and the sampling density; A trajectory fusion model is constructed based on multi-source path information, and the candidate paths are fused to obtain a fused trajectory. The fused trajectory is used to construct expert demonstration data and train a path generation strategy model; The path generation strategy model is optimized using reinforcement learning methods to obtain the final path planning strategy.

2. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 1, characterized in that, The path planning algorithms include graph search-based path planning algorithms and random sampling-based path planning algorithms.

3. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 1, characterized in that, The unified representation processing includes path progress normalization and spatial resampling.

4. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 3, characterized in that, The path progress normalization is achieved through path arc length calculation, and its expression is as follows: in Indicates the progress of the path execution.

5. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 1, characterized in that, The trajectory fusion model adopts a multi-source trajectory fusion method based on Kalman filtering.

6. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 5, characterized in that, During trajectory fusion, different path sources are weighted by observing the noise covariance matrix.

7. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 1, characterized in that, The path generation strategy model is trained using an imitation learning method to learn the mapping relationship between the robot's state and the path target point.

8. The robotic arm path planning method based on multi-source trajectory fusion and learning strategy according to claim 1, characterized in that, By interacting with the environment through reinforcement learning methods, the path generation strategy model is continuously optimized.

9. An electronic device comprising one or more processors and a storage device for storing one or more programs, characterized in that, When the program is executed by the processor, the processor performs the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having an executable program stored thereon, characterized in that, When the program is executed by the processor, the processor performs the method according to any one of claims 1 to 8.

Images