Reproduce “Learning to Recover: Dynamic Reward Shaping with Wheel-Leg Coordination for Fallen Robots” on OpenResearch

About this paper

Boyuan Deng, Luca Rossini, Jin Wang, Weijie Wang, Dimitrios Kanoulas, Nikolaos Tsagarakis

This research presents an asymmetric Proximal Policy Optimization (PPO) framework, incorporating Episode-based Dynamic Reward Shaping and curriculum learning, to achieve autonomous fall recovery for wheeled-legged robots. The approach demonstrates up to a 99.1% recovery success rate and quantifies a 15-26% reduction in joint torque consumption, alongside robust zero-shot transferability across different robot platforms.

Key insights

• The dynamic reward shaping (ED) mechanism effectively resolves the exploration-convergence dilemma in complex, sparse-reward tasks by progressively tightening reward constraints, enabling both broad action space exploration and precise posture control.
• Active wheel-leg coordination significantly improves recovery stability and energy efficiency, demonstrating a 15-26% reduction in average joint torque by leveraging wheel actuation for centroidal adjustments and controlled sliding contacts.
• The developed framework enables robust cross-platform generalization and zero-shot sim-to-real transfer, allowing policies trained in simulation to operate effectively on distinct physical robots without platform-specific tuning.

Summary by alphaXiv.