Sampling-Based Safe Reinforcement Learning

This research proposes Sampling-Based Safe Reinforcement Learning (SBSRL), a model-based RL algorithm that maintains safety throughout the learning process by enforcing constraints across a finite set of dynamics samples. The method approximates an intractable worst-case optimization over uncertain dynamics, enabling practical safety guarantees in continuous domains. It also introduces an exploration strategy that eliminates the need for explicit exploration bonuses. Empirically, SBS,RL achieves safe and efficient exploration in both simulation and real robotic hardware, and extends to deep-ensemble implementations for high-dimensional continuous control.