Reproduce “Safety-Aware Imitation Learning via MPC-Guided Disturbance Injection” on OpenResearch

About this paper

Le Qiu, Yusuf Umut Ciftci, Somil Bansal

The MPC-SafeGIL framework enhances imitation learning policy safety by injecting optimal adversarial disturbances during expert demonstrations, enabling policies to learn robust recovery behaviors in safety-critical scenarios. This method consistently reduced collision rates and improved task success across diverse robotic platforms, achieving, for instance, a 75% real-world safety rate for a Crazyflie quadrotor compared to 25% for baseline methods.

Key insights

• Proactive injection of optimal adversarial disturbances into expert demonstrations can effectively expose safety-critical states, enabling imitation policies to learn robust recovery behaviors.
• Sampling-based Model Predictive Control (MPPI) provides a scalable and practical method for computing these adversarial disturbances, applicable to high-dimensional and black-box systems where analytical models are unavailable.
• Structured, adversarial disturbance injection is more effective than random noise for generating informative training data, leading to policies with superior safety, robustness, and generalization capabilities.

Summary by alphaXiv.