Aditya Shirwatkar

“This paper solves a critical deployment gap for legged robots — RL locomotion policies trained without safety constraints will eventually step on or collide with real-world objects, and this paper proposes a practical, modular filter that intercepts and corrects unsafe footstep plans in real-time without retraining the underlying policy.”

“Our safety filter is agnostic to the internals of π and requires only that the policy accepts contact locations as input.”

“Applying both projections consistently yields the best safety-performance tradeoff across all optimizers... projection was also the most significant component of our proposed algorithmic improvements to reduce the number of safety violations in planning and execution.”

“We deploy the predictive safety filter on a Unitree Go2 hardware, where the proprioceptive state is streamed to an external PC running the optimizer, while the computed contact targets are returned to the robot via a tethered connection.”

“All experiments are run on a single NVIDIA RTX 3090, with rollouts accounting for the bulk of the computation.”