FailGuard: An OoD-Simulation Approach for Proactive Failure Detection in Visuomotor Policy Learning

Leiyu Wang, Jun Lv, Qiao Sun, Yifei Wu, Ao-Bo Wang, Cewu Lu, Qinying Gu, Nanyang Ye

Keywords: Embodied AI, Failure detection

Abstract: Failure detection is a critical capability for ensuring the safety of robotic systems, which can anticipate and avoid irreversible harm to the environment, the robot itself, or to the human being during interacting with the physical world. However, failure detection in real-world robotic manipulation has long been challenged due to the inefficiency and potential risks when collecting diverse failure data through rollouts. In this paper, we introduce a method named \textbf{FailGuard} to achieve failure detection for robot visuomotor manipulation policy learning. To learn a failure detector, we only collect success rollouts from the pre-trained manipulation policy and augment the rollouts to simulate failure cases for detection learning. After trained with the collected data, the failure detector can achieve runtime failure detection and early stop the impending failure, improving the success rate and safety of the robot manipulation. We evaluate our method on both RoboSuite benchmark and real world tasks. The experimental results show that our proposed method outperforms several different kinds of baseline methods, and can effectively prevent failure cases during manipulation. Project Website: \url{https://zuyu3.github.io/failguard.github.io/}

Primary Area: applications to robotics, autonomy, planning

Submission Number: 8539