CONTACT (CONtact-aware TACTile learning for robotic disassembly) asks a simple question with hard consequences: how much better can a robot manipulate when it can feel, not just see?

We extended the Stanford Diffusion Policy framework to ingest GelSight tactile readings alongside camera observations, building an end-to-end multi-modal pipeline for data collection, training, and on-robot deployment. The policy learns from demonstrations and conditions its action prediction on the fused visual–tactile state, which matters most exactly when vision falls short: during occluded contact, slip, and fine insertion.

Trained policies were deployed on physical hardware to verify the approach on tactile-rich disassembly tasks, where the tactile channel produced measurable success-rate improvements over a vision-only baseline. The work has been accepted to IEEE/RSJ IROS 2026, the field’s flagship robotics conference.