RAMPS: Robust Adaptive Multi-Step Predictive Shielding
RAMPS is a scalable predictive shielding framework for safe reinforcement learning in high-dimensional nonlinear systems. It combines learned dynamics, robust multi-step control barrier functions, and minimally invasive action projection.
Research summary
The project addresses safe exploration: an agent must learn high-performing policies while maintaining safety constraints during training. RAMPS models system dynamics in a linear or lifted representation, looks ahead over multiple steps, and adjusts an action only when needed for safety.