Active Tactile Pose and Shape Estimation of Highly Dynamic Objects

Ethan Kroll Gordon, Michael Posa

Keywords: active learning, online learning, tactile sensing, contact modeling, manipluation

TL;DR: This work proposes a marginalization and sampling-based computation of expected information gain amenable to highly dynamic rigid-body contact.

Abstract: Learning a physically accurate object model at test time can provide significant benefits in predictability and reuse between tasks. Tactile sensing compliments vision with its robustness to occlusion, but its temporal sparsity necessitates careful online exploration to maintain data efficiency. Direct contact can also cause an unrestrained object to move, requiring both shape and location estimation. In our main conference work, we introduced an active learning and exploration framework that uses only tactile data to simultaneously determine the shape and location of rigid objects with minimal robot motion. This abstract proposes an extension to that work, addressing the fact that information quantification in dynamic systems requires reasoning about the likelihood of past measurements given a current state. Though marginalization and Markov Chain Monte Carlo (MCMC) trajectory sampling, we can avoid the difficulties of poorly-conditioned ``backwards'' simulation and nearly-discontinuous contact dynamics. The resulting information matrix in principle should be more robust to highly-dynamic trajectories such as tumbling.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 16