How to control robots safely when they touch things

Robots doing contact-rich tasks like pushing or grasping need smooth dynamics models for gradient-based control, but smoothing parameters hide the true contact forces—making safety filters unreliable. This work provides a way to bound actual contact forces despite smoothing, then builds a safety framework that automatically tunes the smoothing level and adds adaptive margins to handle force prediction errors. Simulations show it eliminates violations that standard safety methods miss.