Conceptual

Safe Learning-Based Control for Robotic Systems

How to learn high-performance control policies for real robots while preserving safety properties such as stability, obstacle avoidance, over-actuation prevention and liveness. It places control on a spectrum from classical control theory (differential-equation models, PID, the linear-quadratic regulator, model predictive control, control Lyapunov functions with guarantees) to deep reinforcement learning (model-free policies improved from experience), and examines safe augmentations that combine the two plus embedded-system considerations for real deployments.