PaperlayerView exact revision on arXiv

Source-linked AI summary

SAFE: Unified Slip and Fracture Detection with Low-Cost Acoustic Sensing in Robotic Grasping

Zerun Wang, Vivek Kamat, Shekhar Bhansali

arXiv:2610.08802v1cs.RO

TL;DR

Fragile-object grasping requires detecting both slip and fracture when material properties are unknown. SAFE combines passive acoustic sensors and motor feedback in a unified classifier and adaptive controller. It reports higher success than fixed-force baselines and succeeds on objects absent from training.

  • Problem

    Robots grasping fragile objects need to recognize slip and fracture when object material properties are unknown.

    Source: §I · p. 1

  • Method

    SAFE combines two passive PVDF acoustic sensors and motor proprioception in a unified classifier for normal, slip, and fracture states, integrated with adaptive grasp control.

    Source: §I · p. 2, D. Model Selection and Training · p. 5

  • Results

    91.3% grasp success across five object categories versus 71.7% and 60.9% for fixed-force baselines, with 82.4% success on unseen categories.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    The results show failure-aware grasp control using low-cost passive acoustic sensing without object-specific calibration.

    Source: §V · p. 7

    The training data covered two fracture objects and three slip objects, and some per-condition evaluation counts were modest (n = 4–15).

    Source: §V · p. 7

Abstract

from arXiv · show

Manipulating fragile objects remains challenging as robots must understand the state of what they grasp, such as slip or fracture, to respond appropriately, especially when material properties are unknown. In this paper, we present SAFE: a low-cost, general-purpose sensing approach that detects both slip and fracture in real time using two passive polyvinylidene fluoride (PVDF) acoustic sensors and motor proprioception, without relying on vision or prior material knowledge. The sensors are mounted on a compliant Fin Ray gripper, and a unified HistGradientBoosting classifier reports the state (normal, slip, or fracture) from a 79-dimensional feature vector. Under leave-one-grasp-out cross-validation, SAFE achieves an Alert-F1 of 0.884 with near-zero slip-fracture confusion, and ablations confirm that acoustic sensing is indispensable. An adaptive grasp controller built on this detection layer runs at 104 Hz on a Jetson Orin Nano, achieving 91.3% success across 46 closed-loop robot trials spanning diverse object categories, while each fixed-force strategy drops to 0% on object conditions that mismatch its preset. It further reaches 82.4% success on novel objects unseen during training, demonstrating robust, failure-aware grasp control without object-specific calibration.

I. INTRODUCTION

SAFE combines passive acoustic sensing with motor feedback to distinguish slip, fracture, and normal contact, then uses those detections for adaptive grasping. The paper reports improved closed-loop success over fixed-force baselines and performance on unseen object categories.

I. INTRODUCTION

  • Reliable grasping must address both slip as a lower force boundary and fracture as an upper one; slip-only feedback can risk damage during force increases.Fracture detection identifies an existing crack, allowing force escalation to stop before damage propagates.Source: §I · p. 1
  • Prior acoustic methods focused on slip, leaving open whether one learned sensing system could distinguish slip and fracture for real-time control.Ablations identify acoustic sensing as required for reliable detection.Source: §I · p. 2
  • SAFE uses two passive PVDF films on a Fin Ray finger with motor current and position feedback for unified slip, fracture, and normal-state detection.The 79-dimensional classifier runs at approximately 104 Hz on a Jetson Orin Nano.Source: §I · p. 2
  • Only 3 of 437 alert windows were assigned to the opposite failure mode under leave-one-grasp-out evaluation.The protocol avoids trial-level leakage.Source: §I · p. 2
  • 91.3% grasp success across five object categories, versus 71.7% and 60.9% for fixed-force baselines; unseen categories reached 82.4%.The unseen-object evaluation covered 51 trials across three object categories absent from training.Source: §I · p. 2

II. RELATED WORKS

Prior tactile work supports contact-rich manipulation through force, shape, texture, and slip information, but comprehensive reconstruction is not necessary for every grasp-control task. Recognizing a small set of task-critical events can suffice for closed-loop control.

A. Tactile Sensing for Contact-Rich Manipulation

  • Visual-tactile Fin Ray systems provide contact perception, while imaging rate and camera form factor constrain their temporal responsiveness and structure.Barometric sensing offers low-latency force and slip estimates, but temperature-driven pressure drift can limit use in thermally variable settings.Source: A. Tactile Sensing for Contact-Rich… · p. 2
  • For fragile-object grasping with unknown materials, fracture and slip define opposite boundaries of the safe grasping window.Fracture reflects force exceeding object tolerance; slip reflects insufficient force to hold the object.Source: A. Tactile Sensing for Contact-Rich… · p. 2
  • Many grasp-control tasks can use recognition of a few task-critical interaction events instead of full contact-state estimation.The relevant events depend on the task, rather than requiring comprehensive tactile reconstruction in every case.Source: A. Tactile Sensing for Contact-Rich… · p. 2

B. Fracture Detection

Prior fracture sensing focuses on excessive force, while slip sensing addresses insufficient force; each alone covers only one boundary of safe grasping. The related work motivates acoustic sensing as a possible shared modality for both failure modes.

B. Fracture Detection

  • Fracture research uses pressure or deformation sensing, while acoustic-emission studies report crack-related waves; fracture-only perception still cannot determine sufficient holding force.The cited acoustic-emission literature describes signals emerging before yielding and bursts accompanying crack initiation and growth.Source: B. Fracture Detection · pp. 2–3

C. Slip Detection

  • Slip perception supplies the complementary lower force bound, but without fracture detection, increasing force to address slip risks exceeding the fracture limit.Prior work includes tactile controllers and acoustic methods estimating continuous slip direction and magnitude.Source: C. Slip Detection · p. 3

D. Acoustic Sensing

  • Acoustic signals carry information about both fracture and slip, motivating a shared sensing layer; prior studies had not tested unified recognition of both for closed-loop control.SAFE addresses this gap with two passive PVDF channels and motor proprioception feeding one classifier.Source: D. Acoustic Sensing · p. 3

III. METHODOLOGY

SAFE addresses material-independent, real-time detection with compliant hardware, aligned acoustic and motor signals, and a windowed feature representation. Its two-channel passive sensing hardware is designed to limit stiffness impact while remaining low-cost.

III. METHODOLOGY

  • SAFE targets material-independent detection on low-cost edge hardware by combining compliant sensing, windowed features, and one classifier for two failure signatures.The three stated constraints are no prior material knowledge, real-time operation, and minimal sensing hardware.Source: §III · p. 3

A. Hardware Design

  • Two PVDF films on a compliant Fin Ray gripper capture transient vibrations alongside motor current and position data.The sensors sit on inner and outer surfaces of one finger to minimize stiffness impact; aligned ROS 2 topics support multimodal feature extraction.Source: A. Hardware Design · pp. 3–4
  • Approximately $84 in component cost covers the complete sensing hardware, compared with $355 for DIGIT and $510 for GelSight Mini.The listed hardware includes two PVDF films, a dual-channel charge amplifier, and two USB audio interfaces.Source: A. Hardware Design · p. 4

B. Dataset Construction

The training dataset combines fracture and slip recordings from distinct object sets, yielding 1,499 labeled windows across 252 grasp-trial recordings. Fracture events were labeled using synchronized video, while slip trials varied object weight and gripper current.

B. Dataset Construction

  • The dataset contains 1,499 labeled 500-ms windows across 252 grasp-trial recordings.It includes 293 fracture windows, 144 slip windows, and 1,062 normal windows.Source: B. Dataset Construction · p. 4
  • Fracture data came from eggshells and potato chips, with synchronized video used to label detected events as fracture or normal.The gripper current was gradually increased after contact until fracture or the motor current limit.Source: B. Dataset Construction · p. 4
  • Out-of-distribution testing used onion ring chips, a soap-coated glass beaker, and a plastic cup.The glass beaker and plastic cup were collected with varied weights.Source: Figure 2 · p. 4
  • Slip data came from hand-soap bottles, paper cups, and water bottles across varied weights and gripper currents.Trials used a vertical lift and were labeled slip or no-slip.Source: B. Dataset Construction · p. 4

C. Feature Extraction

SAFE uses complementary time- and frequency-domain signal features to distinguish fracture, slip, and normal contact. Each 500-ms window yields a 79-dimensional feature vector combining acoustic and motor-proprioception information.

C. Feature Extraction

  • Fracture signals show sharp impulses, slip shows continuous higher-amplitude oscillations, and normal contact has a comparatively quiet baseline.Median spectra are visibly separable; fracture has more low-to-mid-frequency energy than normal contact, while slip has the least.Source: C. Feature Extraction · p. 4, Figure 3 · p. 5
  • At deployment, features use a 500-ms buffer; offline event recordings are cropped into one or more independently labeled windows.Fracture events have a median duration of 495 ms.Source: C. Feature Extraction · pp. 4–5
  • Each window produces 79 features: per-channel spectral and time statistics, cross-channel measures, and motor current and position statistics.The feature vector includes spectral shape, time-domain statistics, cross-channel correlation and ratios, and motor-state summaries.Source: C. Feature Extraction · p. 5

D. Model Selection and Training

SAFE trains a three-class HistGradientBoosting classifier on the full feature vector, selecting its configuration with grasp-grouped cross-validation. The adaptive grasp policy uses slip and fracture detections to adjust grip force during approach and after lifting.

D. Model Selection and Training

  • HistGradientBoosting classifies fracture, slip, and normal from all 79 features, using balanced class weights because normal comprises 71% of samples.Sequential decision trees use histogram-binned features for fast training and inference.Source: D. Model Selection and Training · p. 5
  • Hyperparameters were screened with 8-fold GroupKFold and finalists re-evaluated with Leave-One-Grasp-Out cross-validation over 252 grasp-trial groups.The selected model uses depth 5, learning rate 0.1, minimum 20 samples per leaf, and L2 regularization 0.1.Source: D. Model Selection and Training · p. 5
  • The classifier predicts fracture when its confidence reaches 0.5; otherwise, it predicts slip at confidence 0.5 or normal.Fracture confidence is checked before slip confidence.Source: D. Model Selection and Training · p. 5

E. Grasp Control Strategy

  • The adaptive controller repeats test lifts and increases current by 1.2 times after slip until a lift succeeds or the motor limit is reached; detected fracture aborts the trial.During the lift, it decreases current slightly after fracture detection and increases it slightly after slip detection.Source: E. Grasp Control Strategy · p. 5

IV. EXPERIMENTS

Under Leave-One-Grasp-Out evaluation, SAFE achieved an Alert-F1 of 0.884, with few confusions between slip and fracture. Ablations support acoustic sensing, and inference ran at approximately 104 Hz.

A. Classification Performance

  • Alert-F1 was 0.884 [0.856, 0.908], with slip F1 of 0.916 and fracture F1 of 0.851 under Leave-One-Grasp-Out evaluation.The Alert-F1 is the mean of slip and fracture F1; the reported confidence interval uses group-level cluster bootstrap.Source: A. Classification Performance · p. 5, Table II · p. 6
  • Only 3 of 43 misclassified alert windows were assigned to the opposite failure mode; the other 40 were confused with normal.The confusion matrix illustrates the low cross-confusion between slip and fracture.Source: A. Classification Performance · p. 6, Figure 4 · p. 6
  • Mean inference latency was 9.638±0.155 ms per recording, corresponding to approximately 104 Hz across 125 deployment recordings.The recordings comprised 126,368 total inferences on the Jetson Orin Nano.Source: A. Classification Performance · p. 6
  • Adding the outer PVDF channel improved slip F1 by +0.0347 (p = 0.0164), while its fracture-F1 change of +0.0024 (p = 0.8352) was not significant.The ablation comparison was All versus Inner PVDF + Motor.Source: Table III · p. 6, A. Classification Performance · p. 6
  • At this dataset scale, CNN and Random Forest underperformed HGB, and a peak-acoustic-feature decision list scored below HGB on every metric.The unified model’s overall performance was statistically indistinguishable from two separate detectors, which had higher slip F1 but required two inferences per window.Source: Table IV · p. 6, A. Classification Performance · p. 6

B. Closed-Loop Evaluation

The adaptive controller achieves higher overall success than either fixed-force baseline while suppressing both slip and fracture failures. The evaluation compares three controllers across 46 trials spanning five objects and multiple weight conditions.

B. Closed-Loop Evaluation

  • Fixed Low (50 mA), Fixed High (300 mA), and Adaptive were compared across 46 trials spanning five objects and multiple weight conditions.Success required lifting and moving the object without visually observable slippage or damage; failure included fracture, slip, or a model-triggered abort.Source: B. Closed-Loop Evaluation · p. 6
  • 91.3% overall success for Adaptive compares with 71.7% for Fixed Low and 60.9% for Fixed High, suppressing both failure modes without prior knowledge of object fragility.The figure breakdown shows Fixed Low fails primarily by slipping, while Fixed High fails primarily by fracturing.Source: B. Closed-Loop Evaluation · p. 6, Figure 5 · p. 6
  • Fixed Low never fractures, whereas Fixed High never slips; fixed-current grasping works well only when object properties are uniform and known in advance.Whiskers in Fig. 5 show 95% Wilson confidence intervals on success rate.Source: B. Closed-Loop Evaluation · p. 6, Figure 5 · p. 6

C. Out-of-Distribution Generalization

The adaptive controller retains 82.4% aggregate success across three object categories absent from training, without object-specific calibration or retraining. The evaluation includes a soap-coated glass beaker, onion ring chips, and a plastic cup.

C. Out-of-Distribution Generalization

  • The soap-coated glass beaker tests slip detection on both a novel material and an induced low-friction surface absent from training.The tested categories were onion ring chips, the beaker, and a plastic cup.Source: C. Out-of-Distribution Generalization · p. 7
  • 82.4% aggregate success was achieved without object-specific calibration or retraining across the unseen object categories.Success was 77.8% for the beaker, 86.7% for onion rings, and 83.3% for the plastic cup.Source: C. Out-of-Distribution Generalization · p. 7
  • Performance degrades gracefully rather than catastrophically, with more lift tests and inferences needed to reach sufficient gripping current.The beaker achieved 77.8% success, which the authors say suggests generalization beyond material-specific acoustic patterns.Source: C. Out-of-Distribution Generalization · p. 7

V. CONCLUSION

SAFE combines passive acoustic sensing and motor proprioception to detect normal, slip, and fracture states and support adaptive grasp control. It reports strong closed-loop results, while its training data and per-condition trial counts bound the current evaluation.

V. CONCLUSION

  • SAFE uses two low-cost passive PVDF sensors and motor proprioception on a compliant Fin Ray gripper with one HistGradientBoosting classifier for three grasp states.The framework requires neither prior knowledge of object material nor expensive sensing hardware.Source: §V · p. 7
  • A representative soap-coated glass-beaker sequence shows slip detection, repeated current escalation, and a successful lift after the controller reaches a stable grasp.The figure presents one representative test lift of 13 total.Source: Figure 6 · p. 7
  • 0.884 Alert-F1 under Leave-One-Grasp-Out evaluation accompanies 91.3% closed-loop success and 82.4% success on previously unseen objects at approximately 104 Hz.The evaluation reports near-zero confusion between slip and fracture.Source: §V · p. 7
  • PVDF sensors are presented as a plug-and-play sensing layer for existing grippers and grasp-control strategies without substantial hardware redesign.SAFE can also operate as a standalone adaptive grasp controller.Source: §V · p. 7
  • Training spans two fracture objects and three slip objects, while several per-condition evaluation counts are modest at n = 4–15.The authors identify broader material coverage, including ductile or non-brittle failure modes, and more per-object trials as next steps.Source: §V · p. 7
  • Future work includes larger and more diverse datasets, multi-finger PVDF and different grippers, physics-informed learning, severity analysis, and fracture anticipation.The authors also envision tokenizing estimated grasp states for higher-level manipulation planning in future Vision-Language-Action systems.Source: §V · pp. 7–8
Loading 2610.08802v1…