SlipSense: Multimodal Tactile Learning for Low-Latency and Generalized Slip Detection
Tong Jian, Aditya Thurvas Senthil Kumar, Xinyi Li, Ziling Chen, Tianyu Dai, Ali Sengul, Matteo Grimaldi, Wenjie Lu, et al.
cs.RO · cs.AI · cs.LG
Quick verdict
SlipSense fuses pressure and vibration for 96.7% Macro F1 and ~23 ms slip detection with zero-shot hand transfer, but needs custom hardware and hides a slow tail.
Summary
SlipSense is a multimodal tactile slip detector built on TacV5, a fingertip module pairing a 32x32 piezoresistive array (942 active taxels, 240 Hz, normal pressure only, 12 bit) with a 3-axis MEMS accelerometer at 8 kHz. The two streams are encoded separately: a small three-block CNN maps each pressure image to a 128-d embedding, while accelerometer data is high-passed at 10 Hz, converted to a causal 16 ms log-mel spectrogram, aggregated across axes by log-sum-exp for provable direction invariance, and encoded with SSAST-Tiny. Per-sensor embeddings are max-pooled so the model is agnostic to sensor count, then combined by bidirectional cross-attention and a causal temporal layer that predicts no-contact, no-slip or slip at 240 Hz. Ground truth comes from a Mark-10 motorized test stand with 0.02 mm encoder resolution, with slip onset defined at 0.07 mm of crosshead displacement, giving labels independent of the sensor signal. On 1.4M frames from 37 objects (28 train, 9 held out) the fused model reaches 96.7% Macro F1 at 1.33% false-positive rate, versus 81-84% F1 for either modality alone, and detects 76% of events within 23.1 ms. Trained only on a UMI gripper, it transfers zero-shot to a Tesollo dexterous hand (94.39% F1 two-finger, 87.24% three-finger) and prevents object loss in 95 of 100 human pull trials.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 6 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Piezoresistive pressure and accelerometer vibration provide complementary slip cues with complementary failure modes, so cross-attention fusion reaches 96.7% Macro F1 at 1.33% FPR versus 81-84% for either modality alone.
- SlipSense operates at 240 Hz with 76% of slip events detected within 23.1 ms including inference, and 2.3 ms inference on an RTX A4500.
- A model trained only on a UMI parallel-jaw gripper transfers zero-shot to a Tesollo dexterous hand across unseen objects, unseen sensor units and different finger counts without retraining.
- Ground-truth labels from a Mark-10 test stand at 0.02 mm resolution make slip onset measurable independently of the sensor signal, enabling the latency characterization prior work lacks.
- Integrated into a closed-loop re-grasp controller, SlipSense detects all slip events in 100 human pull trials and prevents object loss in 95.
Strengths
Ground-truth labels are independent of the sensor signal, which is rare and important for this task
Section 3.2 uses a Mark-10 F105-EM motorized test stand whose crosshead position is recorded at 0.02 mm resolution, with slip onset defined as cumulative displacement above 0.07 mm. Because the object is non-stretchable and clamped, crosshead displacement maps directly to slip at the gripper surface. This avoids the circularity of prior work that derives onset from optical markers [3, 27], friction criteria on the sensor signal itself [18, 28], or subjective human annotation [5, 21].
The latency and false-positive trade-off is measured jointly, which the authors correctly identify as missing from prior work
Table 1 reports median detection latency in frames alongside FPR, accuracy and Macro F1 for every method. The learned single-modality rows show the trade-off explicitly: Piezo-only gives FPR 7.54 percent but 20.50-frame latency on UMI ID and 36.17 frames on OOD, while XL-only gives 2.14 percent FPR at 4.67 frames. Fusion improves both simultaneously (1.33 percent FPR, 2.00 frames), which is the paper's strongest single result.
Zero-shot cross-platform transfer is tested along three separate axes, not just one
Ablations are run in the fused setting, which is the right control for an encoder comparison
Closed-loop deployment is reported with an honest failure analysis
Weaknesses
Headline latency hides a heavy tail driven by slow slip
No learned baselines: comparisons are analytic only
Reported accuracy is inflated by a dominant no-contact class and no per-class breakdown is given
No dispersion despite a three-seed protocol, and no significance tests
Cross-platform transfer depends critically on a single augmentation choice
Training data covers a single linear slip axis while deployment claims cover rotational slip
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SlipSense argues that two sensing modalities with complementary failure modes should be fused for slip detection, and that the field has been evaluating the wrong things. The scope is well chosen: the authors identify detection latency and cross-platform generalization as the two under-reported axes, build hardware that makes both measurable, and evaluate on 1.4M frames covering 37 objects (28 train, 9 held out) across five test splits with progressively increasing shift. Four claims carry the paper: Piezo and XL are complementary and jointly necessary; the fused model reaches 96.7% Macro F1 with FPR below 1.6%; 76% of events are detected within 23.1 ms; and a model trained purely on a UMI gripper transfers zero-shot to a Tesollo hand. Claims about latency are properly hedged in Section 7, which concedes that only a single linear slip axis appears in training data.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 6 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SlipSense and what hardware does it need?
SlipSense is a machine-learning framework that detects when a grasped object starts to slip, using a custom fingertip sensor called TacV5. TacV5 combines a 32x32 piezoresistive pressure array sampled at 240 Hz with a 3-axis MEMS accelerometer sampled at 8 kHz, synchronized into frames over a CAN-FD link. The pressure array captures how contact force is distributed spatially, while the accelerometer captures the friction-induced vibration that accompanies relative motion. SlipSense encodes each stream separately, fuses them with cross-attention, and outputs no-contact, no-slip or slip at 240 Hz. The sensor is a prototype from Analog Devices and is not a commodity part, so replicating the work requires building the hardware.
How accurate and how fast is SlipSense?
On held-out objects from the same gripper, the fused model reaches 96.98% frame accuracy and 96.77% Macro F1 with a 1.33% false-positive rate, compared with 81-84% Macro F1 for either sensing modality alone. 43% of slip events are detected on the very frame slip begins and 76% within 20.8 ms; adding 2.3 ms of inference on an NVIDIA RTX A4500 gives 23.1 ms end to end. Speed matters a lot: at 18 mm/s over 90% of events are caught within 20.8 ms, but at 2 mm/s the mean delay rises to 14.9 frames (about 62 ms) with a worst case of 106 frames.
Does SlipSense work on robot hands it was not trained on?
Yes, without any retraining. Trained only on a UMI parallel-jaw gripper with two sensor units, it was evaluated zero-shot on a Tesollo DG-5F dexterous hand with physically distinct sensor units: 94.39% Macro F1 at 1.14% false-positive rate in two-finger pinch, and 87.24% at 3.12% in three-finger pinch. Moving the contact from fingertip to fingerpalm on the UMI gives 94.56% Macro F1. The authors note the transfer relies partly on accelerometer augmentation that suppresses platform-specific vibration signatures: without it, Tesollo two-finger false positives rise from 1.14% to 11.31%.
How were slip labels obtained without trusting the sensor itself?
The authors used a Mark-10 F105-EM motorized test stand as an external labeling instrument. The crosshead clamps one end of a non-stretchable object while the sensor-equipped gripper holds the other, so crosshead displacement translates directly into slip at the gripper surface. Position is recorded at 0.02 mm resolution, and slip onset is defined as cumulative displacement exceeding 0.07 mm. This avoids labels derived from optical markers, from the tactile signal itself, or from human judgment, and the Mark-10 is absent at deployment time.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 2, 3, 5, 6, 7, 8, 9, 10, 11, 12 and 13 cannot be inspected; all numbers come from captions, tables and body text, including Figure 7(a) latency percentages and Figure 7(b) window saturation values.
- Appendix sections are included but the paper states a project page is coming soon and no repository, dataset release or hardware documentation is available in the text, so availability claims rest on the authors' statements.
- The definition of false-positive rate does not specify whether the denominator is all frames or only non-slip frames, so cross-paper FPR comparisons should be treated cautiously.
- Deployment results (100/100 detection, 95/100 prevention) are narrated without per-trial measurements in the text, and the four operators are not identified or balanced in any reported breakdown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.