Predictive Safety Filters for Contact Rich Quadruped Locomotion

Aditya Shirwatkar, Aaron M. Johnson, Majid Khadiv, Shishir Kolathaya

Keywords: Legged Robots, Safety Filters, Optimization

TL;DR: A sampling-based safety filter for contact-rich quadruped locomotion that optimizes contact locations via full-physics rollouts to enforce whole-body collision avoidance

Abstract: Learning-based policies have enabled quadruped robots to perform increasingly complex, contact-rich locomotion. However, these policies provide no safety guarantees, making deployment risky when domain shifts introduce unseen obstacles. Even though large-scale offline safe learning can circumvent this, it is not practically feasible to cover all edge cases. Classical safety frameworks like Hamilton-Jacobi Reachability, Control Barrier Functions, and Model Predictive Shielding offer formal guarantees but face fundamental limitations for legged systems: the curse of dimensionality, difficulty constructing valid barrier functions, and overly conservative recovery policies. We propose a minimally invasive predictive safety filter that targets likely safety rather than worst-case guarantees. Our approach formulates a sampling-based receding-horizon optimization over foot contact locations, warm-started by a nominal input and an optional recovery policy, and is bootstrapped with a learned value function. By optimizing in the compact contact-location space, the filter redirects footsteps around obstacles while an underlying contact-conditioned policy trained offline maps contacts to joint-level behavior. We evaluate the filter in simulation on a Unitree Go2 across three scenarios: (i) large-obstacle navigation, (ii) dynamic obstacle avoidance, and (iii) cluttered-environment traversal. Lastly, we show that it substantially reduces safety violations while preserving goal-reaching performance.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 19