Robotics

TacPAC Corrects a Grasp Mid-Chunk Using Predicted Touch

Robb Harlan 4 min read

Most world-action models guess the future, then freeze the plan until the next chunk. TacPAC (arXiv:2609.05266, 4 September 2026) keeps that plan and lets new fingertip images rewrite the part that has not executed yet.

The authors — Zipei Ma, Xiaofei Wei, Junzhe Jiang, Shunlin Lu, and Li Zhang — put the code on GitHub under Logos Robotics Group. The robot in the paper is a Flexiv Rizon 4.

TacPAC teaser comparing failed tight grasps and insertions with successful tactile-corrected trials
Fragile grasp and plug insertion, with and without tactile correction. Source: TacPAC GitHub / arXiv:2609.05266.

Predict, cache, patch

Stage 1 trains a tactile-predictive world-action model: a video expert forecasts future RGB and tactile frames while an action expert denoises a chunk. Stage 2 freezes that base. For each planned chunk, TacPAC caches the predicted contact and the action representation tied to it. A tactile expert then reads each new tactile image against that cache and outputs a delta on the unexecuted suffix.

The README’s timing: one correction is 30.4 ms (32.9 Hz) in their setup, 20.7× faster than regenerating the whole chunk. The paper’s line is the same idea — feedback is interpreted against what the plan expected, not as a raw reflex.

TacPAC two-stage architecture with video expert, action expert, and tactile expert sharing a KV cache
Stage 1 predicts vision and touch; stage 2 corrects the live suffix through a tactile-action cache. Source: TacPAC GitHub.

Five contact-rich tasks

Twenty real-world trials per task. Average success:

VariantPlugFruitChipBottleCardAvg.
Vision only1530605022
Tactile prediction, no expert355065201537
Correction, no prediction402545302533
Both, no cache405075403047
TacPAC806590404564

The tasks are charger-plug insertion, multi-object fruit transfer, potato-chip transfer, empty-bottle uprighting, and expansion-card insertion. The 16-point gap over the strongest ablated baseline is the cache: predicted contact plus the plan’s own keys, reused every tick.

Datasets and checkpoints are “being prepared.” The repo currently has model, training, preprocessing, and deployment code.

A Human’s Take

Predicting touch is cute. Using that prediction as the thing you compare the real fingertip against, while the arm is still moving, is the part I will steal. 22% to 64% on five messy tasks is a real jump. Bottle uprighting still sitting at 40% is the honest number. Ship the datasets.

Sources