Vision-based obstacle avoidance robotic arm path planning based on a multi-level PPO framework
Qi Sun, Jianhao Guo, Guobing Sun
Abstract
In complex and dynamic operational environments, there is a growing demand for intelligent robotic manipulators equipped with autonomous path planning and real-time obstacle avoidance capabilities. In response to the sluggish reaction and limited adaptability of conventional path planning algorithms in dynamic environments, this study develops a multi-level path planning algorithm that integrates visual perception with hierarchical motion planning strategies. The proposed algorithm utilizes YOLOv10 to process RGB-D images and construct dynamic 3D occupancy maps, enabling high-precision obstacle recognition within the scene, with an achieved detection accuracy of 99.8%. Subsequently, by integrating with RRT path search and B-spline trajectory smoothing, and through PPO, an end-to-end perception-decision-execution closed-loop path planning is achieved. Experimental results demonstrate that the algorithm achieves path planning success rates of 100% in obstacle-free scenarios, 97.0% with a single static obstacle, 95.67% with multiple static obstacles, and 89.3% in dynamic obstacle environments. These findings validate the algorithm's robustness and adaptability across various typical settings, highlighting its strong generalization capability and potential for practical deployment.