TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback

Jianbo Zhou

SJTU · SCUT

Boyuan Zhao

ECUST

Yuzheng Zhang

ECNU

Yiyang Chen

SJTU

Wenxin Chen

SJTU

Qiuyue Li

PolyU

Xiangyang Gu

CCUT

Yuhan Cao

UESTC

Xiao Xia

SJTU

Yanzhe Hu

SJTU · HUST

Zhijie Deng ^†

SJTU

Preprint · 2026

^† Corresponding author

Paper Code · Coming soon BibTeX

Abstract

Contact-rich manipulation requires adapting to contact states that can evolve substantially within an action horizon. However, chunk-based vision-language-action models predict complete action chunks from observations collected before execution, leaving tactile conditioning stale during execution. Existing tactile-reactive approaches typically rely on separate high-frequency controllers, which increase both architectural and training complexity. In this paper, we introduce TacForcing, a streaming action-generation framework that effectively incorporates execution-time tactile feedback. Instead of employing a separate reactive controller, TacForcing replaces the standard action expert with a streaming action expert to generate actions conditioned on the evolving tactile observations acquired during execution. TacForcing also introduces Execution-Aware Tactile Attention (EATA), which restricts tactile conditioning to actions nearing execution, thereby reducing the temporal mismatch between tactile acquisition and action execution. Across six simulated UniVTAC tasks and three real-world contact-rich manipulation tasks, TacForcing achieves average success rates of 65% and 69%, respectively, outperforming strong baselines in both settings.

65%

Simulation success

Average over six UniVTAC tasks

69%

Real-world success

Average over three contact-rich tasks

TacForcing framework: a VLM and Streaming Action Expert connected to an execution-time tactile feedback loop — TacForcing progressively completes, executes, and tactilely refines action blocks inside one streaming generation process.

Why execution-time touch?

Vision-language-action policies commonly predict an entire action chunk from observations captured before execution. That works well when the scene changes slowly, but contact states can shift sharply within the same horizon: grasp force changes, slip begins, or a fingertip meets an occluded surface while the visual scene remains nearly unchanged.

Conditioning the complete chunk on one initial tactile frame therefore creates a growing temporal mismatch. TacForcing addresses that mismatch without adding a separate high-frequency reactive controller.

Comparison showing visual observations changing slowly while tactile observations evolve during action execution — Tactile representations can evolve substantially inside a single action chunk even when visual observations remain similar.

Streaming generation with fresh tactile feedback

TacForcing divides an action horizon of $H=KB$ actions into $K$ consecutive blocks, each containing $B$ actions. Instead of assigning one flow time to the entire chunk, block $k$ follows its own schedule:

\lambda_k^{(n)} = \min\left(\frac{n}{kS}, 1\right).

Near-term blocks become ready first. Once a block is executed, the latest fingertip observations are encoded and generation resumes from the retained intermediate states of every unfinished block.

Generate a block

Near-term actions reach clean states first while later blocks retain their intermediate states.

Execute & sense

The ready block is executed, then fresh tactile observations are encoded from the fingertips.

Resume with touch

EATA routes the latest tactile tokens only to the action block scheduled to execute next.

Execution-Aware Tactile Attention

The newest tactile observation is most relevant to the action block that will execute next; applying it to distant blocks may make those actions depend on contact information that will be stale by the time they run. EATA uses a simple visibility mask so that only the next block can directly attend to the current tactile tokens:

\mathcal{M}_{i,m}^{(k)} = \begin{cases} 0, & b(i)=k,\\ -\infty, & \text{otherwise}. \end{cases}

The same temporal structure is used during training and inference. This keeps the model aligned with the tactile information that will actually be available at each execution stage.

Real-world demonstrations

All videos below are encoded and displayed at 5× speed. Front and side views play together when both recordings are available.

Transfer Liquid

The robot manipulates a transparent dropper to draw liquid from a flask and dispense it into a beaker, requiring precise grasp and contact regulation despite partial visual occlusion.

Front view 5× SPEED

Side view 5× SPEED

Transfer Liquid · Blue

Front view 5× SPEED

Side view 5× SPEED

Transfer Liquid · Gold

Wipe Board

The robot moves an eraser across a marked whiteboard while maintaining sufficient surface contact to remove the marks.

Front view 5× SPEED

Side view 5× SPEED

Wipe Board · TacForcing

Front view 5× SPEED

Side view 5× SPEED

Wipe Board · Face

Front view 5× SPEED

Side view 5× SPEED

Wipe Board · Patch

Stand Bottle

The robot grasps a bottle lying horizontally on the table, reorients it in hand, and places it upright while maintaining a stable grasp.

Front view 5× SPEED

Side view 5× SPEED

Stand Bottle

Experiments

Simulation results

TacForcing achieves the best average success rate and the highest or tied-highest score on five of the six tasks.

The six UniVTAC simulation tasks used to evaluate TacForcing — Six contact-rich manipulation tasks from the UniVTAC benchmark.

Method	Lift Bottle	Pull-out Key	Lift Can	Put Bottle in Shelf	Insert Hole	Insert Tube	Avg.
π₀.₅	88	43	46	43	39	48	51
UniVTAC-ACT	59	41	24	6	36	58	37
RDP	84	18	12	41	23	75	42
FTP-1	89	35	66	23	62	76	59
TacForcing	90	48	63	43	69	79	65

Real-world results

TacForcing reaches a 69% average success rate across the three real-world tasks.

Real-world Stand Bottle, Transfer Liquid, and Wipe Board tasks with success-rate comparison — Representative real-world task snapshots and success rates across all three tasks.

What matters? Ablating execution-time touch

A single fixed tactile observation does not consistently improve performance. Streaming the tactile updates helps, and aligning those updates with the next block through EATA yields the strongest results on every evaluated ablation task.

Average success rate

Three simulation tasks / three real-world tasks

Simulation Real world

Base

Simulation

43%

Real world

42%

Fixed tactile

Simulation

42%

Real world

31%

TacForcing w/o EATA

Simulation

51%

Real world

48%

TacForcing FULL MODEL

Simulation

60%

Real world

69%

Adding EATA to streaming tactile updates raises average success from 51% to 60% in simulation and from 48% to 69% in the real world. Relative to fixed tactile conditioning, the complete method gains 18 and 38 percentage points, respectively.

Conclusion

TacForcing turns action generation itself into the feedback loop. It progressively completes action blocks, executes them, refreshes tactile observations, and resumes generation from the unfinished states. EATA then ensures each tactile update affects the actions close enough to execution to use it. The result is a simpler path to tactile reactivity—and consistent gains across both simulated and real contact-rich manipulation.

BibTeX

@misc{zhou2026tacforcing,
  title        = {TacForcing: Streaming Action Generation with
                  Execution-Time Tactile Feedback},
  author       = {Zhou, Jianbo and Zhao, Boyuan and Zhang, Yuzheng and
                  Chen, Yiyang and Chen, Wenxin and Li, Qiuyue and
                  Gu, Xiangyang and Cao, Yuhan and Xia, Xiao and
                  Hu, Yanzhe and Deng, Zhijie},
  year         = {2026},
  eprint       = {2608.25798},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  doi          = {10.48550/arXiv.2608.25798},
  url          = {https://arxiv.org/abs/2608.25798}
}