A robot-control system that refreshes tactile, or touch, feedback as actions unfold reported higher success than the comparison policies in the paper's tests. Called TacForcing, it averaged 65% success across the simulation evaluation and 69% across the real-world evaluation.
The figures come from a preprint and are descriptive success rates, not estimates backed by reported uncertainty intervals or inferential tests. The evaluation counted 100 rollouts for each simulation task and 16 independent trials for each real-world task. Those sample sizes and the absence of reported uncertainty mean the percentage-point gaps should be treated as results from these tested settings, not as a precise measure of performance everywhere.
The problem is timing
TacForcing is aimed at a timing problem in tactile control. When a policy generates a complete action chunk before execution, the tactile conditioning attached to that chunk can become temporally stale. The proposed Streaming Action Expert progressively generates near-term actions while retaining intermediate states for actions that come later.
An additional mechanism, Execution-Aware Tactile Attention (EATA), restricts the current tactile representation to the next scheduled action block. Later blocks receive updated tactile feedback when they become next, linking the newest touch information to the part of the movement about to be executed.
The reported training setup used 50 demonstration trajectories per simulation task for 15,000 training steps, and 100 trajectories per real-world task for 30,000 steps. The global batch size was 256 in both settings.
One analysis illustrates the timing issue, but it is not a broad survey of episodes. In a representative 40-action dropper-squeezing horizon, the final sample came 35 control steps, or 1.17 seconds, after the initial observation. The reported cosine distances - a measure of change between internal representations - were about 0.005 for vision and 0.55 for touch, showing much greater tactile than visual movement in that example.
The strongest gaps appeared in the averages
In simulation, TacForcing was highest or tied-highest on five of six tasks. Lift Can was the exception: its success rate was 63%, compared with 66% for the best-performing method. Across the simulation tasks, the average was 65%, 14 percentage points above π0.5 and 23 points above RDP.
The real-world average was 69%. TacForcing was 17 percentage points above FTP-1, 27 points above GR00T N1.7 and 42 points above π0.5. On Transfer Liquid, it recorded 50% success, while no baseline exceeded 19%.
A closer look at the design
To examine the design's components, the ablation compared four configurations on three simulation tasks and three real-world tasks. The reported configurations were Base, Fixed Tactile, TacForcing without EATA and the full TacForcing setup. Because this comparison covered fewer simulation tasks than the main evaluation, its averages are a separate, narrower test.
Fixed Tactile averaged 42% in simulation versus 43% for Base, and 31% in the real world versus 42% for Base. In the no-EATA streaming comparison, the averages were 51% and 48%, compared with 42% and 31% for Fixed Tactile. With EATA, the full configuration averaged 60% in simulation and 69% in the real world - 18 and 38 percentage points above Fixed Tactile, and 17 and 27 points above Base, respectively.
The component comparisons are consistent with an association between the full design and higher reported averages, but they do not by themselves show that EATA or execution-time updating alone caused the gains. The real-world comparisons rest on 16 independent trials per task.
The evidence remains narrow
The same caution applies to the representation analysis: it was based on one representative episode, not a population-level estimate. More broadly, the evidence is limited to the robot-policy tasks and configurations reported in these comparisons, so it does not establish performance beyond them.
Even with those limits, the preprint raises a practical question for robot control: whether touch is most useful when it is refreshed for the next action rather than held fixed across a longer chunk. TacForcing's reported averages favored the refreshed approach in the tested settings, while Fixed Tactile did not improve the averages.
Paper data and sources
Original title: TacForcing: Streaming Action Generation with Execution-Time Tactile Feedback
Authors: Jianbo Zhou, Boyuan Zhao, Yuzheng Zhang et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text