Tactile Exploration with Particle-Based Belief Entropy

Published: 02 Jul 2024, Last Modified: 15 Jul 2024DM 2024EveryoneRevisionsBibTeXCC BY 4.0
Track: Paper Submission Track
Keywords: Contact-rich Manipulation, Tactile Exploration, Belief Space Control
TL;DR: This paper extends belief space control to continuous, particle-based belief states in the context of contact-rich manipulation under uncertainty with sparse tactile feedback.
Abstract: This paper extends belief space control to continuous, particle-based belief states in the context of contact-rich manipulation under uncertainty with sparse tactile feedback. We answer the open question of how to quantify information gain for a continuous particle-based belief by proposing a new approximation for the entropy of a particle belief that also captures the object-robot interaction dynamics. Moreover, we address the challenge of the discontinuous and sparse nature of the measurement signal by proposing a sampling-based information-gathering controller that selects the next best action from a set of sampled candidate trajectories based on the approximated entropy of predicted future belief states. In robot experiments, we show that action selection based on the approximated particle entropy significantly improves the information-gathering process in terms of efficiency and success rate of a subsequent open-loop grasp.
Supplementary Material: zip
Submission Number: 192
Loading