Reproduce “Simultaneous Extrinsic Contact and In-Hand Pose Estimation via Distributed Tactile Sensing” on OpenResearch

About this paper

Mark Van der Merwe, Kei Ota, Dmitry Berenson, Nima Fazeli, Devesh K. Jha

Researchers from the University of Michigan and MERL developed TacGraph, a factor graph framework that simultaneously estimates a grasped object's in-hand pose and extrinsic contacts using distributed tactile sensing and physical constraints. This approach achieved superior accuracy in tactile-only scenarios and led to significantly higher success rates in a peg insertion task.

Key insights

• Jointly estimating in-hand object pose and extrinsic contacts, while enforcing comprehensive physical constraints, effectively resolves ambiguities inherent in local tactile observations.
• The factor graph framework provides a robust and principled way to fuse diverse sensory information and physical laws, allowing accurate estimation even in the complete absence of visual data.
• Successive contact events can be leveraged to iteratively prune ambiguous pose initializations, leading to a more reliable and physically consistent final estimate.

Summary by alphaXiv.