Failure Prediction from Few Expert Demonstrations

Anjali Parashar, Kunal Garg, Joseph Zhang, Chuchu Fan

Keywords: Failure Detection, Gaussian Process Regression, Bayesian Inference

TL;DR: Three-step methodology for discovering failures that occur in the true system by using a combination of a minimal number of demonstrations of the true system and the failure information processed through sampling-based testing of a model dynamics

Abstract: This extended abstract presents a novel three-step methodology for discovering failures that occur in the true system by using a combination of a minimal number of demonstrations of the true system and the failure information processed through sampling-based testing of a model dynamical system. The proposed methodology comprises a) exhaustive simulations for discovering failures using model dynamics; b) design of initial demonstrations of the true system using Bayesian inference to learn a GPR-based failure predictor; and c) iterative demonstrations of the true system for updating the failure predictor. As a demonstration of the presented methodology, we consider the failure discovery for the task of pushing a T block to a fixed target region with UR3E collaborative robot arm using a diffusion policy and present the preliminary results on failure prediction for the true system.

Submission Number: 87