POLICEd RL: Learning Closed-Loop Robot Control Policies with Provable Satisfaction of Hard Constraints

Proposing POLICEd RL to enforce hard constraints in robotic tasks, significantly outperforming existing methods.