Reference Guided Legged Wheel Locomotion on Stairs with Proximal Policy Optimization

Sammanfattning

Legged-wheel robots combine efficient wheeled motion with the ability toreposition their wheels and body on uneven terrain, but stair climbing remains achallenging contact-rich control problem. This thesis studies reference-guidedreinforcement learning for simulated stair traversal with the LIMX TRON1Awheeled-biped robot. Body and wheel reference trajectories are treated as givengeometric task information, while terrain estimation and online referencegeneration are left outside the scope. A Proximal Policy Optimization (PPO) policy istrained in Isaac Lab using proprioceptive observations and the providedreferences. The experiments evaluate how reward weighting and PPOhyperparameters affect curriculum progression, stability, smoothness, andstair-climbing performance. The results show that reference-guided PPO can learnrobust stair traversal in simulation, including completion of the highestevaluated stair level. However, this behavior only emerges for suitable rewardand optimization settings. A reward balance with stronger three-referencetracking than base-task weighting gave the fastest curriculum progress, whilethe default PPO configuration was the only tested PPO setting that completed allstair levels. The study suggests that structured references are useful forlegged-wheel stair climbing, but must be paired with reward and learning settingsthat make the desired behavior learnable under contact-rich dynamics.