Reinforcement Learning for Robust Climbing Locomotion With Rope-Driven Legged Robot

초록

This study presents a novel control framework for climbing robots that utilizes both rope and leg mechanisms. The proposed robot ascends steep slopes using two ropes while maintaining its balance and adapting its pose to uneven surfaces through its four legs. The robot's overall movement on the slopes is managed by an ascender module, while leg motions are governed by a reinforcement learning (RL) policy trained to sustain local stability under unpredictable disturbances from rope tensions and varying slopes. To enhance stability under partial observability, the policy integrates a latent context vector with a learned Tumble Stability Margin (TSM) for proactive instability detection. Furthermore, to recover from instability in challenging conditions such as slipping or edge hooking, the framework enables dynamic body height adaptation based on stability feedback. Validated via sim-to-real transfer, the system demonstrates that the rope-driven climbing robot maintains consistent locomotion stability across various slope environments and effectively responds to hazardous situations using its learned stability awareness.

키워드