Identifying Precursors to Failures in Robotic Lift-and-Place Tasks

Failure prediction for robotic manipulation in in- dustrial applications has experienced substantial advancements in terms of efficiency and reliability driven by the latest innovations in machine learning. Most of the existing works focus on reactive failure prediction. As an offline analyzing technique, it is not suitable for real-time failure prevention. Therefore, proactive failure prediction becomes a valuable approach to meeting the requirements of online deployment. Although recent research on proactive failure prediction has made progress, it still suffers from limitations such as object specificity, fixed action spaces, and high system complexity. In this paper, we study lift-and- place tasks, and propose a more effective approach to identifying failure by analyzing the relative motion of target objects in the scene. Consequently, we propose a novel method that proactively predict the failure by embedding the relative motions in the scene. We verify our proposed method on both simulation and real- world data. The experimental results indicate that our proposed method not only demonstrates improved generalization but also provides a more precise response to the precursors of failure.

Recommended citation: Zeyu Shangguan, Rajas Chitale, Rutvik Patel, Satyandra K. Gupta, Daniel Seita (2025). "Identifying Precursors to Failures in Robotic Lift-and-Place Tasks; CoRL.
Download Paper