Dynamic Obstacle Avoidance Using Path Reshaping on Probabilistic Roadmaps for High-Degree-of-Freedom Robots

摘要

Robotic manipulators operating in dynamic environments frequently encounter unpredictable obstacles. Existing state-of-the-art motion planning and replanning algorithms struggle to achieve real-time replanning for high-degree-of-freedom robots in these complex, dynamic settings. To address this limitation, we propose a novel local path-reshaping technique to avoid collision from dynamic obstacles in real-time. The developed technique, Probabilistic Roadmap with Deep Neural Network-based Collision-Checking (PRM-DNNCC), leverages a deep neural network (DNN) architecture to accelerate collision checking, which is a critical computational bottleneck during replanning. Upon detecting dynamic obstacles within the workspace, the system locally replans the path by generating a mini-batch PRM around the obstacle. The DNN model is employed to perform both discrete and continuous collision checks on these mini-batch PRMs. A locally optimal path is then executed to ensure safe navigation. The application of the developed algorithm on manipulators with up to six-DOF demonstrates an average success rate of 94.6%, significantly surpassing traditional approaches and proving highly competitive to various other state-of-the-art approaches. Additionally, the DNN-based collision-checking architecture has an accuracy of 98.15% with an average inference time of 6 milliseconds for collision checking, which is faster than all other geometric collision-checking algorithms reported in the literature. We envisage that the developed technique holds promise in motion planning for human-robot collaboration in sectors like healthcare, logistics, manufacturing, and agriculture.