PROPRA anchors pre-touch and tactile signals to robot state and future actions
AIST researchers pretrain sparse fingertip signals against proprioception and upcoming actions. Average real-robot success reaches 75.0%, but the advantage varies by task and remains within seed variation.
Researchers at Japan's National Institute of Advanced Industrial Science and Technology and the CNRS-AIST Joint Robotics Laboratory introduced PROPRA on September 24, 2026. Short for PROprioceptive-and-PRoactive Anchoring, the method pretrains separate encoders for reflective proximity and pressure-sensitive tactile histories by aligning each with the robot's current state and upcoming actions. On four real-robot tasks, it reports a 75.0% average success rate. Paper and version record.
Key takeaways
- Proximity sensing becomes informative during approach, while tactile pressure responds after contact. PROPRA gives both a continuously changing sensorimotor anchor without directly forcing the two sensor embeddings together.
- Across three seeds and 20 trials per seed, PROPRA averages 75.0% versus 71.7% for image-anchored pretraining and 72.1% for image anchoring plus direct tactile-proximity pairing.
- The improvement is not uniform: the image anchor reaches 76.7% on OpenLid versus 73.3% for PROPRA, and code is still labeled “soon.” Official project.
Why naive sensor fusion is not enough
The AIST gripper carries three tactile and six proximity channels. Each encoder receives 0.5 seconds of history. Proximity varies as the fingers approach an object; tactile stays near baseline until contact. After contact, the relationship reverses. In the authors' demonstrations, at least one sensor stays near its stationary value in more than half of the pretraining samples.
Simple concatenation does not consistently help in this experiment. Adding both raw histories to the vision-and-state diffusion policy raises average success from 30.4% to 40.4%, but MovePen falls from 31.7% to 20.0% and OpenLid also underperforms vision alone. This comparison motivates task-specific fusion checks; it does not isolate sparsity as the sole cause of a failure.
PROPRA aligns each sensor encoder independently to an anchor containing the current end-effector pose, gripper opening and next ten subsampled actions, spanning about 1.67 seconds. The contrastive objective never directly pairs tactile and proximity. After pretraining, the sensor encoders are fine-tuned end-to-end with the imitation policy; the anchor and its encoder are removed. The frozen encoders described in the representation probes are a separate evaluation setting.
Results under the reported protocol
The real-robot evaluation uses a UR5e on PickCup, PickSponge, MovePen and OpenLid, with 50 demonstrations per task according to the manuscript’s Section V-A. The project page still says 20 per task; this unresolved source discrepancy should be checked before planning a reproduction. Every policy condition is trained with three seeds and evaluated for 20 trials per seed. PROPRA reaches 81.7% on both pickup tasks, 63.3% on MovePen and 73.3% on OpenLid. The corresponding image-anchor scores are 75.0%, 75.0%, 60.0% and 76.7%. Complete comparison table.
The average lead is 2.9 percentage points over the strongest image-plus-pair variant, or 3.3 points over plain image anchoring, and the paper notes that per-task differences fall within variation across seeds. A fair reading is that PROPRA is competitive and strongest where sensor responses are sparse, not that it wins every manipulation setting.
Representation probes add a more specific result. Over the two seconds before contact, a linear predictor on the frozen proximity embedding achieves the highest time-to-contact R-squared on three of four tasks. On PickSponge it reaches 0.440 versus 0.377 for the image anchor. On MovePen, the image-plus-pair condition is slightly higher, 0.554 versus 0.544.
The authors also collect 100 additional execution episodes per task. After zero-point alignment to compensate for proximity-sensor drift, PROPRA remains positive on every task and seed when a predictor fitted on demonstrations is transferred without refitting. That preprocessing caveat matters: this is not calibration-free deployment.
What this means for robotics
RoboSkin analysis: PROPRA offers a useful alternative to treating camera features as the universal alignment target. A robot's own state and intended action exist throughout the trajectory, including moments when an individual tactile sensor is silent. This can be especially useful for custom fingertips that do not match large visual-tactile foundation-model hardware.
The paper also supplies a warning for multimodal robot learning: an added channel does not help just because it measures contact. The learning objective must tell the model when that channel carries information. Separating pre-touch and post-contact phases may matter as much as increasing sensor resolution.
Limitations and availability
PROPRA is an arXiv v1 preprint evaluated on one gripper, one robot and four grasp-oriented tasks. It does not compare directly with large-scale tactile foundation models, and its pretraining requires recorded robot states and future action segments as anchors. Those future segments are not required at inference. Pretraining takes about three to four hours and policy training about eight hours on one NVIDIA H200 with 141 GB of memory, which is a demanding reference configuration rather than a minimum requirement.
The official project page provides the paper and videos under a CC BY-SA 4.0 page license, but its code link reads “soon.” The November 4 recheck still found Code (soon), without verifying a training repository, model weights, demonstration download or software license. RoboSkin.ai has not reproduced the experiments.
Separate the anchor, sampling recipe and control loop
The proposed training recipe weights selected proximity/contact-transition samples more heavily and adds hard negatives. Image-anchored baselines sample uniformly. The reported comparison therefore changes the sampling recipe as well as the anchor; it cannot assign the entire success difference to proprioception and future actions alone. Sampling and policy integration.
| Quantity | Role | What it does not establish |
|---|---|---|
| Sensor history | Fifty samples from a 100 Hz stream, covering 0.5 seconds | A 100 Hz policy or actuator loop |
| Future action anchor | Recorded training context, removed during deployment | Access to future actions at inference |
| Time-to-contact probe | Regression on frozen proximity features | A deployed collision-avoidance controller |
| Policy success | Three seeds, 20 trials per seed and task | Uniform superiority across tasks or sensors |
The contact threshold used for the probe is in sensor AD units, not a calibrated force limit. Drift alignment remains part of the new-episode analysis. For reuse, log zeroing, sensor-history alignment and the exact sampler; compare anchor choices under a matched sampler if that is the question being tested. Report probe accuracy and task completion separately. These are proposed controls, not experiments performed by RoboSkin. See the training and runtime distinctions.
Sources and related resources
- PROPRA arXiv v1, submitted September 24, 2026
- Full method, tables and limitations
- Official PROPRA project
- RoboSkin tactile foundation models guide
*Source review updated November 4, 2026. Original publication dates are retained. Findings are attributed to their authors; we did not run these systems or verify dataset payloads.*
Continue the topic
