A robot-control system that refreshes its plan with touch information during execution reported the highest average success rates in both simulated and real-world tests. In simulation, TacForcing averaged 65% successful tasks, compared with 59% for FTP-1, 51% for π0.5 and 42% for RDP. In the physical-robot evaluation, it averaged 66%, versus 51% for FTP-1 and T-Rex and 45% for GR00T N1.7.
The results come from an arXiv preprint that tested six UniVTAC simulation tasks and six real-world tasks on two robot platforms. The evaluation used 100 rollouts per simulation task and 16 independent trials per real-world task.
The plan changes while the robot works
TacForcing is built for manipulation tasks in which a robot must make contact with objects. It generates a stream of action blocks, executes the blocks that are ready, keeps unfinished intermediate states, and brings in newly acquired tactile, or touch, feedback as the robot works. Its Execution-Aware Tactile Attention, or EATA, limits the latest tactile information to the next block scheduled for execution.
In the reported implementation, each block contained five actions. The simulation controller used a 50-action horizon split into 10 blocks, while the real-world version used a 40-action horizon split into eight. Completed blocks stayed fixed while unfinished blocks continued through the sampling updates, directing new touch information toward the part of the plan that had not yet been carried out.
During training, the intermediate state of each block was paired with the tactile feedback available at the corresponding execution stage, and the loss was applied only to actions that were still unfinished. The system also used an attention mask that gave the current tactile update direct access only to the next block and excluded the other action positions. In a separate dropper-squeezing diagnostic, the sensor-averaged cosine distance at step 35 was about 0.005 for vision and 0.55 for touch over a 40-action horizon sampled every five control steps. Cosine distance here is a measure of how far a representation has shifted from its starting point.
The advantage was not the same on every task
Against the simulation baselines, TacForcing ranked first or tied for first on five of the six tasks. Its average success rate was 65%, compared with 59% for FTP-1. On Insert Hole and Insert Tube, the reported advantages over FTP-1 were seven and three percentage points, while Lift Can went the other way, at 63% for TacForcing versus 66% for FTP-1.
On physical robots, TacForcing ranked first on five tasks and tied on Wipe Board. Its reported average was 66%, compared with 51% for both FTP-1 and T-Rex and 45% for GR00T N1.7. On three Sharpa North tasks, the reported margin was six percentage points each. TacForcing also scored 50% on Transfer Liquid versus 31% for T-Rex, and 81% on Stand Bottle versus 75% for FTP-1 and T-Rex.
The component tests point to timing
The authors also removed parts of the design to see how the reported averages changed. With EATA disabled, replacing fixed tactile input with execution-time updates changed the average from 42% to 51% in simulation and from 31% to 48% in the real world. Insert Hole showed the largest simulation change, from 34% to 58%.
With execution-time updates retained, adding EATA changed the average from 51% to 60% in simulation and from 48% to 69% in the real world. All six tasks increased in that comparison, with Wipe Board rising from 44% without EATA to 75% with the full system.
A narrow test, not a broad guarantee
These numbers should be read as benchmark comparisons within the tested settings. The evaluation was not randomized, so it does not establish that TacForcing, execution-time updates or EATA caused the higher scores. The ablation results show how the reported averages changed when components were added or removed, but they do not settle whether the same pattern would hold on other tasks or robots.
The scope was six simulation tasks, six real-world tasks and two robot platforms, with 16 independent trials per real-world task. The supplied analysis reports no uncertainty intervals, and the study did not separately assess safety, long-term reliability, force accuracy or generalization beyond the tested settings. That leaves the reported advantage as an encouraging result for the evaluated manipulation benchmarks, not a broad claim about robots in general.
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