SAFE-GIL: SAFEty Guided Imitation Learning for Robotic Systems

Yusuf Umut Ciftci, Darren Chiu, Zeyuan Feng, Gaurav S. Sukhatme, Somil Bansal

Keywords: Data augmentation

TL;DR: A safety-focused behavior cloning approach that prepares robots for high-stakes environments by simulating test-time errors during training

Abstract: Behavior cloning (BC) is a widely used approach in imitation learning, where a robot learns a control policy by observing an expert supervisor. However errors in the learned policy can lead to safety violations, especially in safety-critical settings. While prior works have tried improving a BC policy via additional real or synthetic action labels, adversarial training, or runtime filtering, none of them explicitly focus on reducing the BC policy’s safety violations during training time. We propose SAFE-GIL, a design-time method that improves safety in BC by injecting adversarial disturbances during data collection. These disturbances guide the expert into safety-critical states, better preparing the policy for similar situations at test time. We use a reachability-based formulation to compute these disturbances. Evaluations in ground navigation, aircraft taxiing, and quadrotor flight show that SAFE-GIL significantly reduces safety failures, especially in low-data regimes.

Submission Number: 43