Path Planning for Robotic Arm Bin-Picking Using Improved RRT* Combined with Artificial Potential Field

Introduction

With the advancement of science and technology and the widespread adoption of industrial robots, automation has become a dominant trend in modern manufacturing. Industrial robotic arms are extensively used in tasks such as welding, assembly, and bin-picking, where they must efficiently retrieve parts from bins while avoiding collisions with bin walls and other objects. Path planning is crucial for ensuring smooth and collision-free motion during these operations.

Traditional path planning methods like A* and Dijkstra algorithms are effective for low-dimensional spaces but struggle with high-degree-of-freedom robotic arms. Sampling-based methods, such as the Rapidly-exploring Random Tree (RRT) algorithm, are better suited for high-dimensional state spaces due to their computational efficiency and adaptability. However, RRT suffers from inefficiencies, including excessive random sampling, redundant nodes, and non-optimal paths.

To address these limitations, researchers have proposed various improvements. RRT introduces parent node reselection and rewiring to optimize paths, but it still suffers from slow convergence and large sampling ranges. Informed-RRT restricts sampling to an elliptical region to reduce redundant exploration, yet it remains inefficient when start and goal nodes are far apart. Other enhancements, such as adaptive step sizes and bidirectional search strategies, have been explored, but issues like path smoothness and robustness persist.

This paper presents an improved RRT* algorithm combined with an artificial potential field (APF) for robotic arm bin-picking path planning. The proposed method leverages APF for pre-planning to identify intermediate path nodes that can connect directly to the goal without collisions, reducing unnecessary exploration. Additionally, target-guided sampling, adaptive search parameters, and a pose-based node rejection mechanism enhance efficiency and robustness. Finally, path pruning and smoothing techniques optimize the generated path for smoother execution.

Methodology

Artificial Potential Field Pre-Planning

The artificial potential field method models the workspace as a virtual force field, consisting of attractive forces toward the goal and repulsive forces from obstacles. The attractive force pulls the robot toward the target, while the repulsive force pushes it away from obstacles. The APF is constructed using the target node and obstacle point cloud data.

The pre-planning process involves:

Calculating the attractive and repulsive forces at each point.
Generating a preliminary path by iteratively updating the robot’s position based on the resultant force.
Identifying an intermediate node on the pre-planned path that can connect directly to the goal without collisions.

This intermediate node becomes the new target for the RRT* algorithm, significantly reducing the search space and avoiding unnecessary exploration in obstacle-free regions.

Improved RRT* Algorithm

Target-Guided Sampling

To enhance the algorithm’s directionality, a target-guided sampling strategy is introduced. Instead of purely random sampling, the algorithm occasionally samples the goal node directly, ensuring the tree grows toward the target. The probability of sampling the goal is dynamically adjusted based on the part’s initial orientation, improving efficiency in cluttered environments.

Adaptive Search Parameters

The algorithm adaptively computes the step size and target bias probability based on the obstacle distribution and part orientation. A larger step size is used in sparse obstacle regions for faster exploration, while a smaller step size ensures success in dense obstacle areas. The target bias probability increases when the part is nearly upright, reducing unnecessary exploration.

Pose-Based Node Rejection

Before performing collision checks, the algorithm evaluates the orientation of sampled nodes. Nodes with excessive tilt relative to the initial pose are rejected early, reducing unnecessary inverse kinematics and collision detection computations. This rejection mechanism improves efficiency without compromising success rates.

Path Optimization

Path Pruning

The raw path generated by RRT* often contains redundant nodes. A pruning algorithm removes unnecessary waypoints by checking if intermediate nodes can be skipped without causing collisions. This reduces path length and the number of turns, improving execution efficiency.

Path Smoothing

The pruned path is further smoothed using quasi-uniform cubic B-splines for position interpolation and quaternion spherical interpolation for orientation smoothing. This ensures continuous and smooth motion, minimizing jerky movements during execution.

Experimental Results

Algorithm Comparison

The proposed method was compared with standard RRT and Informed-RRT in two bin-picking scenarios. Results showed:
• Success Rate: The proposed method achieved a 99% success rate, outperforming RRT (81%) and Informed-RRT (92%).

• Planning Time: The improved algorithm reduced planning time by 79.05% compared to RRT and 38.26% compared to Informed-RRT.

• Path Cost: The path cost was reduced by 34.80% compared to RRT and 23.55% compared to Informed-RRT.

Ablation Study

An ablation study confirmed the contributions of each improvement:
• Without APF Pre-Planning: Success rate dropped to 93%, and planning time increased by 140%.

• Without Adaptive Parameters: Planning time increased by 48.96%.

• Without Node Rejection: Planning time increased by 11.37%.

Path Optimization Analysis

Path pruning reduced node count by 65.71% and path cost by 4.36%. Smoothing further enhanced path quality, ensuring stable robotic arm motion.

Simulation and Real-World Testing

Simulations in MATLAB demonstrated smooth joint trajectories without abrupt changes. Real-world tests on a UR3 robotic arm confirmed the method’s practicality, with a total picking time of 5.184 seconds.

Conclusion

This paper presented an improved RRT* algorithm combined with artificial potential field for robotic arm bin-picking path planning. Key contributions include:

APF pre-planning to reduce redundant exploration.
Target-guided sampling and adaptive parameters for efficiency.
Pose-based node rejection to minimize computation.
Path pruning and smoothing for optimal execution.

Experimental results demonstrated significant improvements in success rate, planning time, and path cost compared to existing methods. Future work will focus on singularity avoidance and further optimizing path smoothness and energy efficiency.

doi.org/10.19734/j.issn.1001-3695.2024.08.0310

Was this helpful?

0 / 0