PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yuxiao Zhu, Jinzhou Li, Yifei Dong, Muhammad Suhail, Chunyuan Yang, Xinyuan Luo, Haoyu Li, Boyuan Chen, Xianyi Cheng

arXiv:2609.27381v1cs.RO

TL;DR

Weak contact is difficult to detect from proprioception when contact-induced torque changes are small relative to contact-free variability. CoPRE predicts nominal torque from earlier states and commanded motion, excluding recent observations, then scores torque residuals. On ARX and G1, its tested-trial recall exceeded the reported baselines.

  • Problem

    Weak contact is difficult to distinguish because contact-induced joint torque changes are small relative to contact-free variability.

    Source: §I · p. 1

  • Method

    CoPRE predicts contact-free torque from earlier states and commands, excludes recent observations, and maps torque residuals to a score using noise-weighted Jacobian aggregation.

    Source: §I · p. 1

  • Results

    74.1% recall for CoPRE on ARX and 82.2% on G1 compare with 0% for both baselines on ARX and 16.3% and 42.2% on G1, respectively.

    Source: §I · p. 2

  • Takeaways & Limitations

    ARX manipulation demonstrations provide qualitative evidence of downstream use for contact feedback.

    Source: §I · p. 2

Abstract

from arXiv · show

Contact detection during robotic manipulation allows robots to recognize unexpected contact and adapt their motion accordingly. However, in low-cost robot arms without dedicated force or tactile sensors, detecting weak contacts from proprioception is challenging because the resulting changes in joint-level proprioceptive signals can be small compared to normal variation and noise caused by robot motion itself. We introduce Contact-free Proprioceptive Response Estimation (CoPRE), improving proprioceptive contact detection sensitivity using only contact-free motion, without additional force sensors, contact labels, or analytical dynamics models. CoPRE estimate the expected joint torques under contact-free motion from proprioceptive state history and commanded motion, while removing recent observations that may already reflect contact. It then computes the residual between the expected and observed joint torque estimates, and maps this residual to a contact score using a noise-weighted Jacobian. Real-robot experiments on ARX Arm and Unitree G1 show that CoPRE achieves 74.1% and 82.2% recall on the tested contact trials, compared with 0%/0% on ARX and 16.3%/42.2% on G1 for the learned torque-prediction and inverse-dynamics baselines. CoPRE also reaches 90% detection rate for pushing force at 3.5 N on ARX and 5.5 N on G1. To demonstrate the downstream utility of our method, we implement belief-space manipulation planning for obstacle-aware object placement and book insertion where detected contacts update the spatial belief and enable the robot to retreat from blocked motions, adjust its pose, and retry. Project website at https://copre-arm.github.io

I. INTRODUCTION

CoPRE addresses weak-contact detection on low-cost arms by estimating contact-free torque from proprioception and using residuals to score contact. Real-robot trials report higher recall than the tested baselines, and ARX demonstrations integrate contact feedback into placement and insertion.

I. INTRODUCTION

  • Weak contact is hard to distinguish because contact-induced torque changes are small relative to contact-free variability, while prediction errors vary across joints and motions.Contact-affected states can cause a learned reference to follow the torque change, reducing the residual; raising thresholds to suppress false alarms can miss weak contacts.Source: §I · p. 1
  • CoPRE predicts contact-free joint torque from earlier states and commanded motion, then maps bias-corrected torque residuals to a contact score using noise-weighted Jacobian aggregation.State exclusion withholds recent observations that may reflect contact; contact-free motion alone supports training and calibration, without force sensors, contact labels, or analytical dynamics models.Source: §I · pp. 1–2
  • Belief-guided planning on ARX uses contact decisions, robot pose, and task progress to update a spatial map and guide another placement or insertion attempt.Contact indicates a nearby obstruction, while traversed paths indicate free space; the robot retreats before retrying.Source: §I · p. 1
  • 74.1% recall for CoPRE on ARX and 82.2% on G1 exceed both baselines on each platform; F90 is 3.5 N and 5.5 N, respectively.The corresponding baseline recall values are 0% for both baselines on ARX, and 16.3% for the neural baseline and 42.2% for the dynamics baseline on G1.Source: §I · pp. 1–2
  • The paper implements and evaluates CoPRE on ARX L5 and G1 through baseline comparisons and matched ablations, integrating contact feedback with belief-guided manipulation on ARX.The manipulation demonstrations provide qualitative evidence of downstream use.Source: §I · p. 2

II. RELATED WORK

The related work covers dynamics-based and learning-based contact sensing, alongside uses of contact feedback in manipulation. CoPRE is positioned as learning nominal torque from contact-free data and using contact feedback in ARX and G1 experiments.

A. Model-based contact and force estimation

  • Model-based methods detect contact or estimate external forces from deviations from nominal dynamics, while model errors, friction, and noise limit sensitivity.Hybrid approaches combine semiparametric dynamics learning with disturbance filtering or recurrent momentum-observer uncertainty learning.Source: A. Model-based contact and force… · p. 2

B. Learning-based contact and force sensing

  • Learning-based methods model nominal behavior or learn contact outputs; CoPRE uses no contact or force labels and excludes prediction-interval states.Prior approaches include residual sensing, anomaly detection, and supervised collision, localization, or force prediction.Source: B. Learning-based contact and force… · p. 2

C. Contact sensing and feedback for manipulation

  • Contact feedback supports motion adaptation and tasks including object tracking, active search, retrieval, and crowded bookshelf insertion.The paper evaluates CoPRE on ARX and G1 and uses belief-guided placement and insertion on ARX.Source: C. Contact sensing and feedback for… · p. 2

III. PROBLEM DEFINITION

The problem is to detect weak external contact from built-in proprioceptive signals by distinguishing contact-induced torque changes from contact-free variability. The detector maps observed states and joint commands to a score and binary decision.

III. PROBLEM DEFINITION

  • The experiments target weak contact, defined by small contact-induced joint torque changes relative to contact-free variability, at the end effector or through a grasped object.The detector uses built-in proprioceptive signals; observed torque is the controller’s joint torque estimate obtained from built-in motor signals.Source: §III · p. 2
  • A causal detector maps observed states and joint commands to a nonnegative score and binary contact decision, while the true contact state is unavailable to it.The decision rule requires consecutive score samples at or above a threshold; the goal is high weak-contact sensitivity while limiting false alarms.Source: §III · p. 2

IV. COPRE

CoPRE predicts nominal joint torque from contact-free recordings and converts torque residuals into a contact score. Training, validation, and calibration use separate contact-free sets, with state prediction over a specified horizon.

IV. COPRE

  • CoPRE subtracts predicted nominal torque from observed torque and aggregates bias-corrected residuals through a noise-weighted Jacobian to produce a contact score.A score-history rule converts the score to a contact decision.Source: §IV · p. 2

A. Contact-Free Data, Training, and Calibration

  • Separate contact-free training, validation, and calibration sets fit the predictor, select its checkpoint, and set the detection threshold.Calibration chooses the lowest threshold satisfying the prescribed false-alarm constraint under the deployment confirmation rule; new motions may alter false-alarm rates while the model, statistics, and threshold remain fixed.Source: A. Contact-Free Data, Training, and… · pp. 2–3
  • For a horizon of H steps, the predictor learns the next H states from earlier observations and joint commands.Position and velocity predictions provide auxiliary supervision, while detection uses only the final-step torque prediction; training residuals set the bias and scale.Source: A. Contact-Free Data, Training, and… · p. 2

B. Nominal State Prediction

CoPRE predicts nominal motion without using recent states that may reflect contact, then converts torque residuals into a noise-weighted contact score. The score uses a damped Jacobian fit and consecutive-threshold confirmation.

B. Nominal State Prediction

  • State exclusion predicts H future states from earlier state history and commanded motion, withholding observations over (ta, t] while retaining commands.Detection still uses current torque and joint positions; contact may affect earlier states and commands.Source: B. Nominal State Prediction · p. 3

C. Noise-Weighted Contact Detection

  • Joint residuals are fit by weighted damped least squares, with larger contact-free residual variability reducing a joint’s fitting weight.The fitted vector is an equivalent end-effector force, not a validated contact-force measurement.Source: C. Noise-Weighted Contact Detection · p. 3
  • Damping limits large fitted solutions near poorly conditioned Jacobians, with relative damping λ and numerical floor ϵ.The method uses the resulting weighted Jacobian score as a scalar contact detector; force accuracy and localization remain untested.Source: C. Noise-Weighted Contact Detection · p. 3
  • K consecutive scores at or above γ confirm contact, while a lower score resets the decision and counter.The displayed normalized score is st/γ, with threshold 1.Source: C. Noise-Weighted Contact Detection · p. 3

V. EXPERIMENTS

The experiments examine weak-contact sensitivity, the effects of CoPRE’s design choices, and whether proprioceptive contact detection can guide manipulation without additional force sensing.

V. EXPERIMENTS

  • The evaluation asks whether CoPRE improves weak-contact sensitivity, how its design choices affect sensitivity and false alarms, and how detection can guide manipulation without additional force sensing.Detection questions use benchmarks on both arms; the manipulation question uses separate tasks.Source: §V · p. 3
  • Only CoPRE retains detections under combined contact-free calibration on ARX, while thresholds allow zero confirmed calibration alarms.Fig. 3 curves show threshold-normalized scores at 0.04 m/s; reported percentages aggregate trial-level detection across three pushing speeds.Source: Figure 3 · p. 4

A. Detection Benchmark Setup

The benchmark evaluates contact detection on ARX L5 and the Unitree G1 right arm using built-in proprioceptive signals and shared within-platform evaluation splits. Thresholds are calibrated on contact-free motion, and performance is measured by trial detection, recall, F90, and false-alarm time fraction.

A. Detection Benchmark Setup

  • The benchmark uses a six-degree-of-freedom ARX L5 and the seven-joint right arm of a Unitree G1, with detection based on built-in joint position, velocity, and torque estimates.Neither platform uses dedicated torque sensors; current-derived estimates can be affected by scaling error, temperature, friction, and transmission effects.Source: A. Detection Benchmark Setup · p. 4
  • Each arm pushes five book stacks at three speeds with three repetitions per condition, using separately measured reference sliding resistances from 1.5 to 5.5 N.The reference values characterize each pushing condition and do not track force during individual trials.Source: A. Detection Benchmark Setup · p. 4
  • Contact trials are excluded from fitting and calibration, and methods within each platform share recordings and evaluation splits.Learned models use three seeds, dynamics is evaluated once, and the default H = 3 horizon was selected separately from reported contact trials.Source: A. Detection Benchmark Setup · p. 4
  • The benchmark reports detection within platform-specific evaluation windows, pooled recall across resistances and speeds, F90, and contact-free false-alarm time fraction.F90 is the lowest reference sliding resistance with at least 90% detection at that and all higher tested levels; force at detection is unmeasured.Source: A. Detection Benchmark Setup · p. 4
  • The neural torque-prediction baseline uses joint position, velocity, and tracking-error history, while inverse dynamics subtracts nominal modeled torques before fitting residuals.The neural detector uses an unweighted torque-residual norm; the dynamics detector uses unweighted damped least squares with the translational Jacobian.Source: A. Detection Benchmark Setup · pp. 4–5
  • ARX selects the lowest threshold with zero confirmed calibration false alarms, whereas G1 uses three-fold cross-fitting and a threshold allowing at most 1% false-alarm time in each calibration motion set.G1 holds out task-matched and broad contact-free runs for false-alarm evaluation; its 1% budget was selected during development from a sweep.Source: A. Detection Benchmark Setup · p. 5

B. Detection across Reference Sliding Resistances (Q1)

CoPRE retains contact sensitivity under combined calibration on both robots, while both baselines miss all ARX contacts and neither baseline reaches F90 on G1. Broad contact-free calibration has little effect on CoPRE’s ARX threshold or recall.

Detection performance

  • At 0.04 m/s, Fig. 3 plots threshold-normalized detector scores, with threshold 1, and shows the middle-speed curves rather than rates pooled across speeds.Curves show means and 10–90% ranges across three trials and, for learned methods, three seeds; vertical lines are approximate contact references.Source: B. Detection across Reference Sliding… · p. 5
  • 74.1% ARX recall and 82.2% G1 recall for CoPRE exceed both baselines; F90 is 3.5 N on ARX and 5.5 N on G1.Both ARX baselines miss all contacts, and neither G1 baseline attains F90 within the tested range.Source: B. Detection across Reference Sliding… · pp. 5–6

Calibration coverage

  • Adding broad contact-free calibration raises the neural baseline threshold 2.48–2.63 times and eliminates its detections, while CoPRE’s threshold changes by ≤7.4%.CoPRE recall is 74.8% with task-matched calibration and 74.1% with combined calibration.Source: B. Detection across Reference Sliding… · p. 5

C. Ablation Study (Q2)

The ablations support state exclusion and noise-weighted Jacobian scoring as complementary contributors to sensitivity, with detection outcomes also depending on evaluation window, prediction horizon, and predictor design.

State exclusion and residual aggregation

  • On ARX, state exclusion raises task-matched recall by 34.8 percentage points with weighted Jacobian scoring and 32.6 points with standardized residual norm; on G1, it improves recall with either score.Under combined calibration on G1, weighted-Jacobian recall changes from 83.0% to 82.2%, while standardized-norm recall falls from 49.6% to 31.9%.Source: C. Ablation Study (Q2) · p. 6

Aggregation and noise weighting

  • Noise weighting improves sensitivity on both robots while increasing broad false-alarm time.The comparison holds predictors, kinematics, residual centering, relative damping, and calibration rules fixed.Source: C. Ablation Study (Q2) · p. 6

Evaluation-window sensitivity

  • Shortening G1’s evaluation window from 0.50 to 0.10 s lowers CoPRE recall from 82.2% to 20.7%.State exclusion improves recall at every tested duration.Source: C. Ablation Study (Q2) · p. 6

Prediction design

  • With H = 4 on G1, F90 falls from 5.5 to 4.5 N while broad FAt rises from 0.86% to 1.23%.Shorter horizons reduce recall on both robots; longer horizons retain ARX F90 and raise FAt.Source: C. Ablation Study (Q2) · p. 6

Predictor families

  • The transformer matches the MLP’s ARX F90 with much lower FAt, while the GRU improves both G1 sensitivity and FAt.The transformer is used on both robots to keep the default architecture consistent.Source: C. Ablation Study (Q2) · p. 7

Prediction and temporal filtering

  • Direct multi-step prediction improves recall over autoregressive rollout on both robots; temporal filtering has less consistent benefits and can trade sensitivity for false-alarm time.Relaxing ARX’s calibration FAt budget from 0% to 1% raises recall from 74.1% to 95.6%, while held-out broad FAt rises from 3.32% to 12.41%.Source: C. Ablation Study (Q2) · p. 7

D. Contact-Guided Manipulation (Q3)

On ARX, belief-guided planning uses contact and traversed paths to update a spatial map, then replans blocked placement or insertion attempts. Demonstrations show retreat or pose adjustment followed by a retry.

Belief-guided task policy

  • After each attempt, contact, robot pose, and task progress update a spatial belief map, while traversed paths provide evidence of free space.Contact raises the estimated probability of a nearby obstruction; the planner ranks retries by feasibility, distance to goal, and previous failures.Source: D. Contact-Guided Manipulation (Q3) · p. 7

Protective reaction during placement

  • When a cup blocks placement, CoPRE retreats and places the block elsewhere, leaving the cup upright; the neural baseline continues and tips it over.Task-success rates remain unquantified.Source: D. Contact-Guided Manipulation (Q3) · p. 7

Contact-guided book insertion

VI. DISCUSSION AND CONCLUSION

CoPRE combines state exclusion with noise-weighted Jacobian aggregation and reaches F90 at 3.5 N on ARX and 5.5 N on G1, while neither baseline meets that criterion. The demonstrations are qualitative, and the method’s training, calibration, and evaluation scope impose boundaries.

Conclusion

  • CoPRE reaches F90 = 3.5 N on ARX and 5.5 N on G1; neither baseline meets the criterion within the tested range.Ablations support state exclusion and noise-weighted Jacobian aggregation; ARX demonstrations show protective retreat and book-insertion retries.Source: §VI · p. 7

Limitations

  • CoPRE requires robot-specific contact-free training and calibration, and evaluation covers contacts transmitted through the end effector.Sensitivity to data quantity, motion diversity, unseen motions, and payload changes remains untested; false-alarm rates may change in new conditions.Source: §VI · p. 7
  • During sustained contact, affected states can enter the input history, allowing the reference to follow torque changes and reduce the residual.Future work will test other arm links and quantify detection latency, task success, and false interruptions.Source: §VI · p. 7

Related papers

Loading 2609.27381v1…