CoPRE: Improving Sensitivity in Proprioceptive Contact Detection
for Low-Cost Robot Arms

Project Video · 3 min Experimental Results

Proprioceptive contact detection and reactive manipulation.Download · 22.1 MB ↗

Overview

Weak contact induces changes in joint torque estimates that can be difficult to distinguish from variability due to friction, robot motion, and measurement noise. Contact-free Proprioceptive Response Estimation (CoPRE) predicts nominal joint torque from proprioceptive histories and joint commands, and detects contact through deviations between observed and predicted torque estimates. Proprioceptive observations comprise joint positions, velocities, and motor-derived torque estimates.

CoPRE combines state exclusion within the prediction interval with noise-weighted Jacobian aggregation of torque residuals. Training and threshold calibration use only contact-free recordings, without additional force sensors or contact or force labels. We evaluate contact detection on ARX L5 and the Unitree G1 right arm, and demonstrate contact-guided protective retreat and insertion retries on ARX.

Method

State exclusion limits contact-induced changes to the nominal reference; noise-weighted Jacobian aggregation accounts for contact-free residual variability.

Contact-Free Training and Calibration
Train the nominal-response predictor on contact-free proprioceptive observations and joint commands. Calibrate the detection threshold using separate contact-free recordings.
Nominal Torque Prediction with State Exclusion
Predict nominal joint torque from state–command history preceding the prediction interval and known commands within it. Exclude state observations within this interval to limit the influence of recent contact-induced state changes on the nominal reference.
Noise-Weighted Jacobian Aggregation
Correct the nominal bias in observed-minus-predicted torque residuals, then aggregate them using the end-effector Jacobian with lower weights for joints with greater contact-free residual variability. Confirm contact when the score meets or exceeds the calibrated threshold for three consecutive samples.

CoPRE pipeline: contact-free collection and calibration, a Transformer predictor excluding state observations within the prediction interval, torque residuals and contact confirmation, followed by protective reactions or contact-guided planning. — CoPRE pipeline: nominal-response prediction, torque-residual aggregation, and contact-guided manipulation. View full-size diagram

Experimental Results

We evaluate CoPRE on ARX L5 and the Unitree G1 right arm using 90 physical contact trials. Each platform is tested at five reference sliding-resistance levels and three pushing speeds, with three repetitions per condition (45 trials per platform).

Contact detection under combined calibration. Recall and false-alarm time are reported in percent.
Platform	Method	Recall ↑	F₉₀ (N) ↓	FA time ↓
ARX L5	CoPRE	74.1	3.5	3.32
ARX L5	NEXT-based neural predictor	0.0	>5.5	14.51
ARX L5	Nominal inverse-dynamics detector	0.0	>5.5	0.00
G1 right arm	CoPRE	82.2	5.5	0.86
G1 right arm	NEXT-based neural predictor	16.3	>5.5	1.51
G1 right arm	Nominal inverse-dynamics detector	42.2	>5.5	0.36

CoPRE recall: ARX 74.1% (95% CI: 62.2–85.2%); G1 82.2% (95% CI: 72.6–91.1%). FA time is the fraction of held-out broad contact-free motion spent in confirmed alarm. F₉₀ is the lowest tested reference sliding resistance with ≥90% recall at that and all higher tested levels; >5.5 indicates that the criterion is unmet.

Combined calibration; shared comparison rules within each robot. ARX evaluates detection through push end; G1 uses a 0.5 s window. Reference resistance is measured separately at sliding onset; the 90% criterion must hold at that and all higher tested levels.

Evaluation protocol and metric definitions

ARX selects the lowest threshold with zero confirmed calibration alarms. G1 permits at most 1% alarm time in each calibration motion set and uses three-fold cross-fitting. Training and calibration are separate for each robot.

Learned-model recall averages three seeds; dynamics is evaluated once. The 95% confidence intervals bootstrap physical trials, grouping seed evaluations. The evaluation window begins at a geometric contact reference, not independently measured contact onset.

The broad-motion false-alarm time fraction is the fraction of held-out broad contact-free motion spent in confirmed alarm. Reference resistance characterizes the pushing condition, not force at detection. CoPRE’s score is used for contact detection, not validated force measurement.

ARX book-pushing results across reference resistances of 1.5 to 5.5 newtons. CoPRE retains detections under combined calibration; both baselines lose detections. — ARX scores at 0.04 m/s, normalized by the detection threshold. Blue uses combined calibration; orange uses task-matched calibration. Percentages pool all three speeds.View full-size results ↗

G1 right-arm evaluation across five book stacks, with separately measured reference sliding resistances from 1.5 to 5.5 N. — G1: the same five reference-resistance conditions, with separate predictor training and calibration.

G1 detection rates by reference resistance

G1 detection by reference sliding resistance · 9 physical trials per level
Method	1.5 N	2.5 N	3.5 N	4.5 N	5.5 N
CoPRE	55.6%	85.2%	85.2%	85.2%	100.0%
NEXT-based neural predictor	0.0%	18.5%	0.0%	25.9%	37.0%
Nominal inverse-dynamics detector	22.2%	22.2%	55.6%	44.4%	66.7%

Reactive Manipulation

Qualitative ARX demonstrations connect contact detection to protective retreat and belief-guided retries.

Protective retreat during placement

When a cup blocks placement of a carried block, detected contact triggers a stop and retreat. The robot places the block at an alternate location, leaving the cup upright.

Contact-guided book insertion

Contact with a neighboring book blocks the first attempt. Contact feedback updates a spatial belief map and guides pose adjustments and retries until the book enters the gap.

Two ARX demonstrations: protective retreat and alternate placement beside a cup, and pose adjustment followed by a successful book insertion retry. — Qualitative placement and insertion sequences on ARX.View full-size sequences ↗

Supplementary Experiment Videos

Extended recordings at original speed · silent

CoPRE · protective retreat and alternative placement

The continuous placement sequence shows approach, retreat, and placement at an alternate location.

0:41 · 720p · 1×Download clip ↓

Selected recordings illustrate the experiments and individual manipulation outcomes. Quantitative comparisons are reported in the results above.

Ablation Studies

Matched ablations examine prediction, residual scoring, and temporal design on the same benchmark recordings.

State exclusion improves recall on both platforms. Noise-weighted Jacobian aggregation improves recall relative to unweighted aggregation, with increased false-alarm time. The comparisons below also examine prediction horizons, predictor families, temporal filtering, and detection windows.

Recall and the held-out broad-motion false-alarm time fraction are reported together, in percent, to show the sensitivity–false-alarm trade-off. FA time denotes the fraction of contact-free motion spent in confirmed alarm.

State exclusionARX + G1

State exclusion · combined calibration
Variant	ARX recall ↑	ARX FA time ↓	G1 recall ↑	G1 FA time ↓
Off	3.7	0.42	64.4	0.84
On (default)	74.1	3.32	82.2	0.86

Residual aggregationARX + G1

Residual aggregation · combined calibration
Variant	ARX recall ↑	ARX FA time ↓	G1 recall ↑	G1 FA time ↓
Standardized norm	0.0	11.79	31.9	1.02
Unweighted Jacobian	45.9	1.67	12.6	0.38
Weighted Jacobian (default)	74.1	3.32	82.2	0.86

Prediction horizonARX + G1

Prediction horizon · combined calibration
Variant	ARX recall ↑	ARX FA time ↓	G1 recall ↑	G1 FA time ↓
H = 1	4.4	0.22	70.4	0.88
H = 2	46.7	1.26	71.1	0.94
H = 3 (default)	74.1	3.32	82.2	0.86
H = 4	74.1	4.55	83.7	1.23
H = 5	80.0	7.92	81.5	0.95

Predictor familyARX + G1

Predictor family · combined calibration
Variant	ARX recall ↑	ARX FA time ↓	G1 recall ↑	G1 FA time ↓
Persistence	0.0	0.02	24.4	0.59
Ridge	31.1	0.23	48.9	0.91
MLP	77.8	14.18	81.5	0.84
GRU	57.0	0.72	92.6	0.45
Transformer (default)	74.1	3.32	82.2	0.86

The Transformer architecture is used on both platforms, with separate training for each robot. GRU achieves higher recall and lower broad-motion alarm time on G1.

Prediction and temporal filteringARX + G1

Prediction and temporal filtering · combined calibration
Variant	ARX recall ↑	ARX FA time ↓	G1 recall ↑	G1 FA time ↓
Direct / instantaneous (default)	74.1	3.32	82.2	0.86
Autoregressive	69.6	6.30	65.9	0.81
Fast transient	40.0	1.00	54.8	0.99
Slow sustained	80.0	26.19	50.4	0.56

State exclusion across horizonsG1

G1 · state exclusion off / on · combined calibration
H	Off: recall ↑	Off: alarm time ↓	On: recall ↑	On: alarm time ↓
2	68.9	0.85	71.1	0.94
3	64.4	0.84	82.2	0.86
4	63.0	0.81	83.7	1.23

Detection evaluation windowG1

G1 · state exclusion off / on · combined calibration
Window (s)	Off: recall ↑	On: recall ↑
0.1	13.3	20.7
0.25	38.5	57.0
0.5	64.4	82.2

Predictors and thresholds are fixed across these evaluation windows. The main G1 benchmark uses 0.50 s.

The detection archive contains 45 contact and 42 contact-free ARX recordings, and 45 contact and 51 contact-free G1 recordings. Seeds, cross-fitting, and ablations reuse these recordings. Downloads contain summary tables and a contact-trial index.

Main results CSV ↓Ablation tables CSV ↓G1 temporal checks CSV ↓90-trial index CSV ↓Data notes ↗

Resources

Video and evaluation summaries are available below.

Paper: Not yet released
Research code: Not yet released
Video: Watch · 3 min ↗
Experiment videos: Browse 15 clips ↗
Evaluation data: Ablations & downloads ↗

CoPRE: Improving Sensitivity in Proprioceptive Contact Detection for Low-Cost Robot Arms