Visual imitation policies can plan overall task motion, but contact-rich manipulation often requires gripper adjustments beyond simple open and close commands. The authors of a new arXiv paper propose treating fine-grained gripper control as a local feedback problem that runs within an ongoing task. Their system, ContactPolicy, is a learned controller that takes over the gripper while a visual base policy continues to control the arm.
ContactPolicy combines tactile and force histories with the current robot state to jointly predict gripper increments and their modulation, updating the gripper target from each new observation. An Intent policy manages the handover between the visual base policy and ContactPolicy, deciding when tactile feedback should assume control. In real-world experiments on container opening, object handling, and tool operation, the approach improved task success over both a Visuomotor Policy and a Tactile-Force Policy baseline.
Execution analysis showed how the gripper adjusts contact during a task, including periods of target retention and fine-grained changes as interaction conditions evolve. Arm-state ablations and failure analysis further examined how execution context supports local regulation. The results support the idea of task-local tactile policies for fine-grained gripper control within visual manipulation systems.