TacForcing Brings Execution-Time Touch Feedback to Robot Policies

Introduction

For robots performing grasping, insertion, pressing, and other contact-rich tasks, vision does not reveal everything that matters. A camera may show where an object is, but it may not reliably indicate whether contact is stable, whether the applied force is appropriate, or whether the end effector has encountered an unexpected obstruction. Tactile sensing can provide these signals, yet incorporating them into a vision-language-action (VLA) policy during execution remains difficult.

The synchronization problem

Many VLA systems use action chunks: given observations collected before execution, the policy predicts a complete sequence of actions. Chunking can make inference and control more efficient, but it also creates a temporal mismatch. Once execution begins, the contact state may change substantially, while the planned actions still reflect an earlier observation and cannot immediately react to new tactile evidence.

A common response is to add a separate high-frequency reactive controller. Such a controller can correct actions quickly, but it also introduces another architectural component, additional training requirements, and interfaces between the high-level policy and the low-level controller. The result may be a more capable system, but also a more complicated one.

What TacForcing changes

TacForcing addresses the issue by replacing the standard action expert with a streaming action expert rather than attaching a separate reactive controller. The policy progressively generates and executes action blocks. Actions that have not yet been executed can be refined when new tactile observations arrive, allowing the plan to evolve with the physical interaction.

The paper’s second key component is Execution-Aware Tactile Attention, or EATA. Its design emphasizes three ideas:

• tactile conditioning is concentrated on actions that are near execution;
• information from a tactile observation is less likely to influence actions that are too far in the future;
• the policy can focus its conditioning capacity on the part of the trajectory where fresh contact information is most useful.

This is more specific than simply appending tactile data to the model input. EATA also considers the temporal question of which action segment should be affected by a given tactile observation.

Results and implications

Across six simulated UniVTAC tasks and three real-world contact-rich manipulation tasks, TacForcing reports average success rates of 65% and 69%, respectively, outperforming strong baselines in both settings. The supplied material does not include task-level results or ablation details, so these figures should be read as aggregate performance indicators rather than evidence of uniform gains in every scenario.

The broader contribution is a shift in where tactile feedback is integrated. Instead of treating touch as an add-on for a low-level controller, TacForcing places execution-time tactile information inside the main action-generation loop. If the approach scales to more tasks and sensor configurations, it could offer a simpler route to policies that jointly generate, execute, and revise actions. Important practical questions remain, including inference latency, tactile hardware transfer, streaming computation cost, and reproducibility once the promised code and full implementation details become available.

Source: Hugging Face Daily Papers