Tactile World Models
Overview
Tactile world models study how robots can predict contact, deformation, slip, and other tactile outcomes of their actions, often alongside future visual observations. Common approaches learn action-conditioned dynamics in compact tactile representations, combine touch and vision with contact-aware or asymmetric attention, and use forecasts for planning, corrective control, or policy training. Work spans simulated and real-robot benchmarks for tasks such as insertion, delicate grasping, and in-hand manipulation, with models increasingly designed for varied sensors and high-frequency control. A central open tension is that better tactile prediction does not always improve task success, while systems must balance useful contact information with reliable visual prediction, efficient inference, and generalization across tasks and hardware.
Latest papers
30 papers6 papers in the last four weeks, down 14% on the four weeks before. 0.0% of all new papers.