TLDR: Robot controllers are increasingly complex to support harder and more diverse task requirements, but unintended side effects frequently emerge, potentially leading to costly failures. How to comprehensively test a controller and gain a holistic understanding prior to deployment? Use RoCUS! It elicits instances highlighting a specified behavior, which can then be inspected for systematic patterns or analyzed using interpretability tools.

Abstract: As robots are deployed in complex situations, engineers and end users must develop a holistic understanding of their behaviors, capabilities, and limitations. Some behaviors are directly optimized by the objective function; typically these include success rate, completion time or energy consumption. Other behaviors — e.g., collision avoidance, trajectory smoothness or motion legibility — are typically emergent but equally important for safe and trustworthy deployment. Designing an objective which optimizes every aspect of robot behavior is hard, so we instead propose a method for systematic analysis of exhibited behaviors. RoCUS is a Bayesian sampling-based method to find situations where the robot controller exhibits user-specified behaviors, for example, highly jerky motions. We use RoCUS to analyze three controller classes (deep learning models, rapidly exploring random trees and dynamical system formulations) on two domains (2D navigation and a 7 degree-of-freedom arm reaching), and uncover insights to further our understanding of these controllers and ultimately improve their designs.

Related Work

• Bayes-TrEx: Our prior work that proposed the notion of transparency-by-example and its Bayesian sampling formulation, from which RoCUS adapts to be compatible with practical requirements of robotic evaluations.