Safety-Aware Imitation Learning via MPC-Guided Disturbance Injection

2025-08-05

AI Summary

Imitation learning often suffers from safety violations due to distributional shift, limiting its deployment in safety-critical applications. To address this, we propose MPC-SafeGIL—a novel imitation learning framework that proactively embeds safety into the data generation process. During expert demonstration collection, it injects worst-case adversarial disturbances synthesized via sampling-based model predictive control (MPC), eliminating reliance on analytical dynamics models or real-time expert supervision. The method unifies adversarial robustness, safety-aware dataset construction, and end-to-end imitation learning. Evaluated on high-dimensional black-box systems—including quadrupedal locomotion and vision-based navigation in simulation, as well as real-world quadrotor flight—we demonstrate substantial improvements in policy safety and task success rate. To our knowledge, MPC-SafeGIL establishes the first perturbation-injection-based, safety-driven imitation learning paradigm.

Technology Category

• Machine Learning: Imitation Learning & Inverse Reinforcement Learning
• Multiagent Systems: Adversarial Agents
• Natural Language Processing: Safety and Robustness

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Security and Privacy: Security and privacy of machine learning and AI applications
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

Imitation Learning has provided a promising approach to learning complex robot behaviors from expert demonstrations. However, learned policies can make errors that lead to safety violations, which limits their deployment in safety-critical applications. We propose MPC-SafeGIL, a design-time approach that enhances the safety of imitation learning by injecting adversarial disturbances during expert demonstrations. This exposes the expert to a broader range of safety-critical scenarios and allows the imitation policy to learn robust recovery behaviors. Our method uses sampling-based Model Predictive Control (MPC) to approximate worst-case disturbances, making it scalable to high-dimensional and black-box dynamical systems. In contrast to prior work that relies on analytical models or interactive experts, MPC-SafeGIL integrates safety considerations directly into data collection. We validate our approach through extensive simulations including quadruped locomotion and visuomotor navigation and real-world experiments on a quadrotor, demonstrating improvements in both safety and task performance. See our website here: https://leqiu2003.github.io/MPCSafeGIL/

Problem

• Enhances safety in imitation learning via adversarial disturbances
• Uses MPC to approximate worst-case disturbances for robustness
• Integrates safety directly into data collection for critical scenarios

Innovation

• MPC-guided adversarial disturbance injection
• Sampling-based MPC for worst-case disturbances
• Safety-enhanced imitation learning data collection