A tactile reflex becomes the teacher for fragile grasping
Better Curriculum uses tactile feedback to shape demonstrations for RGB-and-proprioception policies. Its optional runtime gripper arbiter remains a separate source of disturbance correction.
Ziyan Feng and colleagues introduced a different role for touch in a preprint submitted on September 22, 2026: use tactile feedback to shape demonstrations, then train a policy that does not require tactile input at inference. A deterministic 25 Hz controller called TactileReflex produced stable grasps of a thin plastic cup for both Action Chunking with Transformers (ACT) and pi0.5 students. The study also shows the boundary of that idea: a tactile-free student did not retain the controller's disturbance rejection. Paper and version record.
Key takeaways
- With 30 reflex-shaped demonstrations, ACT achieved 19 stable grasps in 20 nominal trials. ACT trained on 30 visually screened manual demonstrations achieved 3 in 20.
- The same data intervention produced 19 of 20 stable grasps with pi0.5, versus 1 of 20 for manual data. These are small physical trial sets on one fragile-cup setup, not a general manipulation benchmark.
- Under randomized lateral disturbances, the same reflex-data pi0.5 policy retained 11 of 20 grasps alone and 20 of 20 with a separate deployment-time tactile arbiter. Disturbance experiment.
What changed
The bottleneck is placed at collection time. Manual gripper commands are delayed and coarse relative to the narrow contact range of a 3.5-gram plastic cup with a 0.3-millimeter wall. The Piper arm uses two MC-Tac vision-based tactile sensors on its gripper. During demonstration collection, TactileReflex adjusts the gripper at 25 Hz from a vision-tactile force proxy while the human controls the arm.
The resulting trajectory still trains a tactile-free student. The ACT model runs at 15 Hz; the two-billion-parameter pi0.5 model uses LoRA and generates actions at 20 Hz, with policy queries observed at roughly 1.7 to 1.8 Hz. The reported training and inference hardware includes one RTX 5090. System description.
Results and the important comparison
The main ACT comparison uses 30 reflex demonstrations and 30 manual demonstrations selected by watching 100 manual trials. Outcomes were labeled stable, loose or dropped from video by five authors. In 20 nominal trials per policy, reflex data yielded 95% stable grasps, while manual data yielded 15%. Adding the tactile arbiter to the manual-data ACT policy raised stable grasps to 50%, still below the reflex-data result.
Could better screening explain the gap? The authors logged tactile contact for a fresh pool of 100 manual demonstrations, ranked them by contact quality and trained ACT once on the top 30. That policy recorded 6 stable grasps in 20 trials, versus 19 of 20 for reflex data. This is a useful control, but it is one training run and still does not randomize the order of data collection.
The unseen paper-cup test is exploratory: reflex-data pi0.5 reached 8 stable grasps in 10 trials and manual-data pi0.5 reached 3 in 10. The paper reports p=0.0698, so the result does not meet the study's conventional significance threshold.
What this means for robotics
RoboSkin analysis: touch can improve a policy using RGB and proprioception without becoming another runtime modality of that learned policy. That matters when production hardware cannot carry the training sensor, or when a fast controller can generate cleaner contact behavior than a human operator. It also reframes robot demonstration data: the controller that creates the trajectory can be as important as model size.
The disturbance result prevents a stronger conclusion. A tactile-free policy may respond through its available visual and proprioceptive observations, but it does not provide the same measured tactile correction during an executing action chunk. Contact-critical systems may still need a low-latency loop, whether from tactile sensors, motor current or a force-torque sensor. Here the runtime arbiter controls gripper behavior; it is not evidence of a validated whole-arm safety system.
Limitations and availability
This is an arXiv v1 preprint and RoboSkin.ai has not reproduced it. The tactile controller signal is a dimensionless vision-tactile proxy, not a calibrated force measurement in Newtons. The approximately 0.5 N deformation observation describes the cup, not the controller's internal value. Collection order was not randomized, and the three outcome labels come from video rather than force ground truth.
The official project page exposes the paper, diagrams and videos. The November 4 review still finds an author’s general GitHub profile rather than a verified dedicated code, dataset or model download. No asset license was established from those reviewed materials. Reproduction therefore requires details and artifacts beyond the visible project page.
Track who controls the gripper at each stage
“Curriculum” here means changing the distribution of demonstrated behavior, not a schedule that presents progressively harder training examples. The robot logs the follower’s realized seven-dimensional action trajectory after the reflex has shaped it. That distinction matters: storing only the human’s raw command would omit the correction the student is supposed to learn. The same operator sees tactile feedback in both manual and reflex-assisted collection. Collection protocol and appendices.
| Stage | Role of touch | Separate acceptance question |
|---|---|---|
| Demonstration collection | A deterministic reflex changes gripper behavior; realized actions are logged | Are corrections and observations aligned, and are all rejected demonstrations counted? |
| Learned policy | ACT or pi0.5 consumes RGB and proprioception | Does the policy retain the demonstrated behavior on held-out trials? |
| Optional runtime arbiter | A tactile loop interprets close/open intent and controls the gripper | Does it release correctly and retain contact during tested disturbances? |
The 30 selected manual demonstrations out of 100 are a curation choice, not a 30% collection-success estimate. Similarly, 20/20 retention in the main disturbance run is not a zero-failure guarantee: four additional training seeds report 79 retained grasps out of 80 with the arbiter, versus 42/80 without it. Keep those repeat runs distinct from the headline single-seed comparison.
Nominal 25 Hz capture does not establish fresh camera input at every control step. Record observation timestamps, policy-query times, action chunks and gripper overrides when adapting the pipeline. Contact-quality occupancy is a frame-level proxy, not calibrated newtons or a complete-trial success rate. These checks connect the method to the manipulation decision map, without treating demonstration quality as a replacement for runtime feedback.
Sources
*Source review updated November 4, 2026. Original publication dates are retained. Results belong to the cited authors; RoboSkin has not run these systems or downloaded their experimental payloads.*
Continue the topic
