Meta-Reinforcement Learning of Hierarchical Fault-Tolerant Controller for Multiple Leg Failures in Hexapod robots

Hexapod robots used in resource exploration and post-disaster rescue operations often face the risk of multiple unexpected leg failures, which can lead to an immediate threat of falling. Traditional controllers generally lack the fault-tolerance required to adjust locomotion gaits to continue to move in a target direction. Maintaining the original motion patterns in such situations can result in severe damage to the robot. This work proposes a hierarchical fault-tolerant control scheme that employs a meta-learning training architecture capable of addressing the non-stationary multi-task leg failure adaptation problem. The meta Learning architecture simultaneously trains multiple locomotion tasks with all possible configurations of one, two or three leg failures occurring during the normal movement of hexapod robots. Experimental results demonstrate the exceptional fault-tolerant capabilities of the trained controller in adapting to unexpected leg failures. Furthermore, the controller maintains the robot’s planar mobility, providing superior velocity tracking and directional following abilities compared to baselines.