Reproduce “SafeSteps: Learning Safer Footstep Planning Policies for Legged Robots via Model-Based Priors” on OpenResearch

About this paper

Shafeef Omar, Lorenzo Amatucci, Victor Barasuol, Giulio Turrisi, Claudio Semini

Researchers at IIT's Dynamic Legged Systems Laboratory developed "SafeSteps," a footstep planning framework that integrates model-based safety priors into a reinforcement learning policy for quadrupedal robots. This approach significantly reduces safety violations by up to 100 times and improves velocity tracking performance by approximately 20% compared to baseline methods, enabling safer and more robust locomotion in challenging environments.

Key insights

• Explicitly integrating model-based safety priors via action masking dramatically reduces safety violations, making the RL policy inherently safer from the outset of training.
• Outsourcing safety constraints to a dedicated perception module (SaFE-Net) allows the RL agent to focus its learning on optimizing dynamic performance, leading to improved tracking and disturbance rejection.
• Combining fast, learned safety classification with reinforcement learning for planning and MPC for actuation creates a robust hybrid system that leverages the strengths of both data-driven and model-based approaches.

Summary by alphaXiv.