Approximating Safety Feedback Without a Safety Oracle via Model Predictive Control
Researchers have developed a new method to approximate safety feedback for robotic control systems without needing a direct safety oracle. This approach leverages simulators to create a proxy for safety functions, bypassing the need for manual labeling or explicit constraint formulation. The algorithm uses Model-Predictive Path Integral (MPPI) to ensure reversibility and verify actions, projecting them to future states and checking if a path back to a previous state exists, thus confirming the state is outside the unsafe set. AI
IMPACT This research could enable safer deployment of robots in complex, unmodeled environments by reducing reliance on explicit safety constraints.