Inverse reinforcement learning (IRL) can extract generalizable behavior from expert demonstrations, but its typical reliance on unconstrained exploration makes it risky for real-world systems. Control barrier functions (CBFs) offer formal safety guarantees, yet designing them analytically is often labor-intensive and highly specialized. This paper proposes a joint solution: restrict the reward function candidacy during IRL to the space of CBFs, so that the learned policy is both safe and continuously improving through experience.
A key result is that the barrier function can be recovered directly from unlabeled expert observations, without requiring explicit safety labels or hand-crafted constraints. The recovered barrier function generalizes to unsafe states that were entirely absent from the expert data, suggesting that the method captures underlying safety structure rather than merely memorizing the demonstrations.
In simulated navigation benchmarks, the approach achieved better safety performance than standard IRL baselines. The authors also compared planning-based and policy-based IRL variants in both simulation and a real-world obstacle avoidance task, shedding light on the trade-offs between the two formulations. The work is a step toward making imitation learning practical for safety-critical robotics applications.