Uncertainty Aware Deployment of Pre-trained Task Conditioned Imitation Learning Policies

Bo Wu, Bruce D Lee, Bernadette Bucher, Nikolai Matni

Keywords: Imitation Learning, Uncertainty Quantification, Generalist Robotics

TL;DR: We suggest a novel way to deploy uncertainty aware policies for pre-trained agents

Abstract: Large scale robotic policies that are trained to perform diverse tasks on many robotic platforms hold great promise; however, reliable generalization remains a major challenge. To address this challenge, it is crucial to appropriately calibrate these models and use the calibrated model to make uncertainty aware decisions across large and diverse sets of data. In this work, we propose an approach to achieve this in pre-trained language conditioned imitation learning agents. Our simulation results using the proposed approach on the pre-trained Perceiver-Actor model demonstrate its effectiveness at improving task completion rates. The code is available at: https://github.com/BobWu1998/uncertainty_quant_peract.git

Submission Number: 15