Foot-End Global Trajectory Planning via GCN-Based Heuristic Tree Search for Crossing Obstacles

Achieving smooth motion for multi-legged robots on complex terrains is a significant focus of research. When encountering high obstacles, robots often need to alter their motion direction to avoid them, increasing redundancy in their motion trajectories. To address this challenge, this paper proposes a method for planning the foot-end trajectory during the swing phase while considering obstacle avoidance without modifying the fuselage trajectory. The method combines the Global Optimal Path Search Tree (GOPST) algorithm and a prior path estimation method utilizing Graph Convolutional Network (GCN). The GOPST explores multiple global paths by conducting local path tree searches in each step, guided by an objective function. To enhance efficiency, redundant search branches with high intensity or high possibility of collision with obstacles or other feet, etc. are eliminated using an event-triggering mechanism based on expert constraints. Another objective function is also formulated to obtain an optimal path that offers a larger safety space and a shorter path length. The optimal path nodes and their environmental features are integrated into a GCN for training. Before the operation of the GOPST, the GCN network provides a preliminary path with fast estimation speed. If the estimated path falls outside the safety margin, the GOPST is reactivated to explore a reliable path. Numerical simulation results validate that the GOPST-GCN approach can rapidly generate a smooth trajectory within the safety space of the foot-end workspace. Furthermore, the search time for finding the optimal path in an untrained environment decreases as the number of tests increases. Experimental verification confirms that robots successfully avoid obstacles by employing foot-end swinging without altering the initial motion direction of the fuselage.