Contact-rich manipulation remains a major challenge in robotics, as robots must adapt to rapidly changing contact states during task execution. Traditional vision-language-action (VLA) models predict entire action chunks before execution, leaving tactile conditioning stale. Existing reactive approaches often require separate high-frequency controllers, adding complexity.
To address this, researchers from Hugging Face introduce TacForcing, a streaming action-generation framework that incorporates execution-time tactile feedback directly into the action generation process. Instead of a separate controller, TacForcing uses a streaming action expert to generate actions conditioned on evolving tactile observations, and an Execution-Aware Tactile Attention (EATA) mechanism to align tactile signals with imminent actions.
In experiments across six simulated UniVTAC tasks and three real-world contact-rich manipulation tasks, TacForcing achieved average success rates of 65% and 69%, respectively, outperforming strong baselines in both settings. The framework progressively generates and executes action blocks while refining unfinished actions with fresh tactile feedback.
The project page, real-world demos, and paper are now available. Code is expected to be released soon.