Online Fault Detection in Manipulation Tasks Via Generative Models

This paper introduces a method, Generative Ad- versarial Networks for Detecting Erroneous Results (GAN- DER), leveraging Generative Adversarial Networks to provide online error detection in manipulation tasks for autonomous robot systems. GANDER relies on mapping input images of a trained task to a learned manifold that contains only positive task executions and outcomes. When reconstructed through this manifold, the input images from successful task executions will remain largely unchanged, while the images from a failed task will change significantly. Using this insight, GANDER enables inspection and task outcome verification capabilities using a large number of positive examples but only a small set of negative examples, thus increasing the applicability of autonomous robot systems. We detail the design of GANDER and provide results of a proof-of-concept system, establishing its efficacy in an autonomous inspection, maintenance, and repair task. GANDER produces favorable results compared to baseline approaches and is capable of correctly identifying off-nominal behavior with 91.65% accuracy in our test task. Ablation studies were also performed to quantify the amount of data ultimately needed for this approach to succeed.