CoPRE detects weak robot contact from proprioception without force labels

Researchers from Duke University, KTH Royal Institute of Technology and Carnegie Mellon University released Contact-free Proprioceptive Response Estimation, or CoPRE, on September 23, 2026. The arXiv v1 method trains on contact-free joint motion, predicts the torque response that should occur without contact and converts the residual into a contact score. It requires neither dedicated force sensors nor labeled collisions. Across 45 physical contact trials on each of two arms, CoPRE reported 74.1% recall on an ARX L5 and 82.2% on a Unitree G1 right arm. Paper and version record.

Key takeaways

• CoPRE excludes recent measured states from its prediction window so that contact does not immediately contaminate the nominal reference, then weights joint residuals by contact-free noise and the robot Jacobian.
• Under each platform's reported calibration protocol, the detector reached at least 90% detection at separately measured sliding-resistance references of 3.5 N on ARX and 5.5 N on G1. These are not force measurements at the instant of detection.
• The official page exposes videos and summary CSV files, but says those tables are manuscript transcriptions rather than raw logs. Placement and insertion demonstrations have no reported task-success rate. Evidence documentation.

What changed

Low-cost arms often estimate joint torque from motor current rather than dedicated torque sensors. A learned nominal model can flag contact when observed torque diverges from expectation, but a model that consumes very recent state may adapt to the collision itself and suppress the residual.

CoPRE addresses that leakage by predicting three steps ahead from an earlier state-and-command history. A 96-dimensional transformer estimates contact-free joint position, velocity and torque. Detection uses the final predicted torque, centers and scales its error with training residuals, and maps the result into Cartesian space with a noise-weighted Jacobian. Three consecutive threshold crossings confirm contact.

Training, checkpoint selection and threshold calibration all use separate contact-free recordings. The ARX windows use typical 32-millisecond intervals and the G1 windows 20 milliseconds. The method still needs a robot kinematic model for Jacobian scoring, even though it avoids a full analytical dynamics model.

What the 90 physical trials show

Each arm pushed five book stacks at three speeds, with three repetitions for every resistance condition: 5 by 3 by 3 equals 45 physical contact trials per arm. The stacks had separately measured peak sliding resistances from 1.5 to 5.5 N. Those reference values characterize test conditions; the researchers did not track force during individual pushes.

On ARX, CoPRE recorded 74.1% recall, while both a learned torque-prediction baseline and a nominal inverse-dynamics baseline recorded zero under the combined calibration setting. CoPRE reached the study's F90 criterion at 3.5 N. Its held-out broad-motion confirmed-alarm time was 3.32%, versus 14.51% for the neural baseline and zero for dynamics. Full evaluation table.

On G1, CoPRE recorded 82.2% recall, compared with 16.3% for the neural baseline and 42.2% for dynamics. Its F90 was 5.5 N, and broad-motion alarm time was 0.86%. The paper's 95% confidence intervals were 62.2% to 85.2% on ARX and 72.6% to 91.1% on G1. Official result summary.

Calibration is part of the result. A retrospective ARX sweep increased recall from 74.1% to 95.6% when the allowed calibration alarm budget rose from zero to 1%, while held-out broad alarm time climbed from 3.32% to 12.41%. This is not a free sensitivity gain.

What this means for robotics

RoboSkin analysis: CoPRE shows how proprioception can provide a useful contact channel when adding tactile sensors is impractical. A contact alarm can be used by a controller to stop motion, update a spatial belief and choose another attempt; the controller must define and validate that response. That makes the method relevant to robot safety and contact-aware manipulation, especially on platforms that already expose joint state and motor-current-derived torque estimates.

It is not a substitute for calibrated force measurement. The detector produces a contact score and confirmed alarm. Its fitted equivalent-force vector is a scoring proxy, not validated force magnitude or localization, and it does not provide a pressure map. Its strongest engineering value may be as a protective or task-state signal layered under a policy, not as complete tactile perception.

Limitations and availability

This is an arXiv v1 preprint and RoboSkin.ai has not reproduced it. The detector needs robot-specific training and calibration. Generalization to new payloads, motion distributions and contacts on other arm links was not tested. Sustained contact may enter the history and reduce the residual. The two contact-guided tasks are qualitative demonstrations, so the paper does not establish a numerical task-success improvement.

The official CoPRE page provides project videos, result summaries, ablation tables and a path-free index of the 90 recordings. Its own documentation says the CSV tables are presentation summaries, not raw logs, and that model seeds reuse physical trials. The September 24 availability check found no implementation link. By November 4, the page links a public CoPRE implementation at e6449cd, with training, calibration, replay and streaming interfaces under a root MIT license. Its reference tests use synthetic signals and analytically assigned model parameters, not released physical recordings or pretrained robot weights. Dependency and robot-model licenses remain separate. We inspected the files without installing or executing them.

A detection window is not a measured stopping delay

With the same G1 predictor and thresholds, recall falls from 82.2% in a 0.50-second evaluation window to 20.7% in a 0.10-second window. The clock begins at a geometric contact-plane reference, not an independently measured physical onset. The result therefore cannot be converted into a guaranteed reaction time from the 20 ms sample interval. Evaluation-window ablation.

For integration, separate three decisions. First, choose and document the motion distribution used for contact-free calibration. Second, validate the alarm with a held-out contact protocol and explicit deadline. Third, test the application response—including stop, retreat, latch and restart—because a detected event alone does not execute those actions.

The public package makes two relevant distinctions in its pinned documentation: an alarm can clear during continued contact, and the application owns the motion response. Its numerical-verification notes also state that generic pooled-time calibration does not automatically impose the paper's separate motion-family budgets or cross-fitting. Running a default example would therefore not reproduce the published experiment.

Record readiness, score, threshold, confirmation, reference event and executed response separately. Use the mechanics-and-measurement comparison when deciding whether an end-effector needs direct tactile sensing as well as arm-level contact alarms. This is a proposed integration check, not a certified protective system.

Sources

• arXiv v1 record, submitted September 23, 2026
• Full CoPRE v1 paper
• Official project and experiment page
• Official notes on the downloadable evidence files

*Source review updated November 4, 2026. Original publication dates are retained. Research results are author-reported; RoboSkin has not built or run these systems.*