TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback
TacForcing lets robot policies update actions with tactile feedback while the action is being executed.
The paper targets contact-rich manipulation, where touch conditions can change inside a predicted action chunk. Its streaming action expert conditions later actions on tactile observations gathered during execution, without adding a separate reactive controller. A second component, Execution-Aware Tactile Attention, limits touch conditioning to actions close to being run. The authors report 65% average success across six simulated UniVTAC tasks and 69% across three real-world manipulation tasks, above their baselines. HF Daily Papers' note
The paper targets contact-rich manipulation, where touch conditions can change inside a predicted action chunk. Its streaming action expert conditions later actions on tactile observations gathered during execution, without adding a separate reactive controller. A second component, Execution-Aware Tactile Attention, limits touch conditioning to actions close to being run. The authors report 65% average success across six simulated UniVTAC tasks and 69% across three real-world manipulation tasks, above their baselines. HF Daily Papers' note
score 4