TL;DR — Using sensed contact to adapt sampling-based robot control online.
ContactKernels uses estimated contact location and object geometry to reshape the proposal distribution of a sampling-based MPC controller during contact-rich insertion.
TL;DR — Building uncertain geometry from vision and contact.
ContactFusion combines visual observations with contact measurements to update a probabilistic geometric representation, allowing interaction to reveal structure that vision alone cannot observe.
TL;DR — Using contact to reduce uncertainty about object pose.
BayesContact maintains a multimodal belief over object pose and uses vision and physical contact to eliminate hypotheses that are inconsistent with interaction.