Uncertainty-Aware Failure Detection for Imitation Learning Robot Policies

Chen Xu, Tony Khuong Nguyen, Patrick Miller, Robert Lee, Paarth Shah, Rares Andrei Ambrus, Haruki Nishimura, Masha Itkina

Keywords: Failure/OOD detection, runtime monitoring, imitation learning

Abstract: Recent years have witnessed impressive robotic manipulation systems driven by advances in imitation learning and generative modeling, such as diffusion- and flow-based approaches. However, these systems can still fail due to suboptimality, inconsistency of stochastic actions, or unfavorable out-of-distribution operating conditions. To ensure dependable operation of such systems before deployment in safety-critical situations, such as in human environments, we need to reliably detect their failures in real-time during inference. In this paper, we propose a modular two-stage approach for failure detection in imitation learning-based robotic manipulation tasks. Our method combines extracting scalar signals that correlate with policy failures and conformal prediction to accurately identify failures while providing statistical guarantees. We investigate both learned and posthoc scalar signal candidates, finding learned signals to be most performant for failure detection. We show the effectiveness of our approach through extensive experiments on diverse robotic manipulation tasks, showcasing its ability to detect failures accurately and quickly. Our results highlight the potential of our method to enhance the safety and reliability of imitation learning-based robotic systems as they continue to improve and become ready for real-world deployment.

Submission Number: 27