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, Saleh Nabi, Tao Yu

Dexterous AI Group, Analog Devices, Inc.

SlipSense fuses high-density pressure (Piezo, 240 Hz) with high-bandwidth vibration (XL, 8 kHz) for accurate, low-latency slip detection that transfers zero-shot across robotic platforms.

UMI Translational Slip

UMI Rotational Slip

Free Motion (No False Alarms)

96.7%

Macro F1 with FPR below 1.6%

23.1 ms

76% of slips detected within

Zero-shot

UMI to Tesollo transfer
No retraining required

SlipSense at a glance: TacV5 sensor on UMI and Tesollo, multimodal signals, frame-level slip prediction — *Left:* TacV5 on UMI gripper and Tesollo hand. *Right:* Piezo pressure frames and XL accelerometer signals are fused by SlipSense for frame-level slip prediction. **Detection latency** is defined as the delay from slip onset to first correct detection.

Summary Video

Abstract

Slip detection is fundamental to dexterous manipulation, yet existing systems often lack precise detection latency characterization and cross-platform generalization. We present SlipSense, a multimodal tactile slip detection framework built on TacV5, a compact sensor integrating a 32×32 piezoresistive array (240 Hz) and a 3-axis MEMS accelerometer (8 kHz). The piezoresistive array captures spatial pressure distribution while the accelerometer captures friction-induced vibration, providing complementary slip cues. The framework performs modality-specific encoding, intra-sensor fusion, and cross-modal attention with causal temporal prediction at 240 Hz.

Experiments on a 1.4M-frame dataset spanning 37 objects demonstrate that the two modalities are crucially complementary: it achieves 96.7% Macro F1 with false-positive rate below 1.6%, detecting 76% of slip events within 23.1 ms. Remarkably, trained purely on a UMI, SlipSense generalizes zero-shot to a Tesollo dexterous hand, transferring across unseen objects, distinct sensor units and new robotic platforms without any retraining.

System & Data Collection

TacV5 Multimodal Tactile Sensor

TacV5 integrates a high-density 32×32 piezoresistive array (Piezo) with a 3-axis MEMS accelerometer (XL) in a compact fingertip form factor. Piezo captures spatial normal pressure distribution at 240 Hz with ~1.55 taxels/mm², while XL captures friction-induced vibration at 8 kHz. We evaluate on two platforms: a customized UMI parallel-jaw gripper and a Tesollo DG-5F 20-DoF dexterous hand.

Data Collection

A single trial proceeds through six phases: approach, grasp, grasp variation, slip, stabilization, and release. The Piezo heatmaps show spatial pressure evolution while the XL spectrogram reveals friction-induced vibration transients at slip onset. Together they provide complementary cues that neither modality captures alone. We collect a 1.4M-frame dataset spanning 37 objects.

Full grasp-slip-detect cycle with Piezo heatmaps and XL spectrogram — Visualization of a complete trial: Piezo pressure maps (top) and XL spectrogram (bottom) across the grasp-slip cycle. The two modalities are complementary: Piezo provides strong contact grounding while XL reacts faster to slip onset.

37 objects spanning cables, connectors, materials, and tools. 28 objects are used for training; 9 held-out objects (yellow boxes) are reserved for evaluation, including Shore hardness test blocks with unseen hardness levels.

SlipSense Architecture

Piezo frames and XL spectrograms are encoded by dedicated backbones, pooled across sensors for variable-cardinality input, fused via cross-modal attention, and passed through a causal temporal classifier that predicts ternary state (no-contact, no-slip, slip) at every frame.

Key Findings

Piezo + XL Are Crucially Complementary

Piezo alone — false alarms from grasp pressure changes (displacement without slip)

XL alone — false triggers from motion vibration without contact

Piezo + XL — Piezo confirms contact, XL confirms relative motion → both failure modes eliminated

Complementarity visualization: Label, Piezo-only, XL-only, and Piezo+XL predictions across two episodes — Frame-level predictions. Red: Piezo false alarms and missed detections. Purple: XL false no-contact reaction (left) and missing initial slip onset (right). Fusion (bottom row) is correct in both cases.

Multimodal Fusion. Fusing Piezo and XL achieves the lowest latency and FPR while substantially improving Macro F1 over either modality alone.
Method	Piezo	XL	UMI ID				UMI OOD
Method	Piezo	XL	Lat (fr)↓	FPR (%)↓	Acc (%)↑	F1 (%)↑	Lat (fr)↓	FPR (%)↓	Acc (%)↑	F1 (%)↑
Baselines	✓		1.00	21.69	72.87	74.51	2.00	21.08	73.84	75.29
		✓	1.00	22.98	79.41	81.79	2.00	19.31	82.99	84.55
	✓	✓	1.00	9.28	81.16	79.96	2.00	7.40	83.33	82.18
SlipSense	✓		20.50	7.54	84.63	83.54	36.17	8.43	82.69	81.91
		✓	4.67	2.14	84.21	81.61	3.00	3.24	84.74	81.33
	✓	✓	2.00	1.33	96.98	96.77	1.50	1.57	95.99	95.75

Higher Density, Better Detection

Higher spatial density captures finer contact patterns that lower resolutions miss entirely.

Piezo resolution ablation: 8x8 (83.2%), 16x16 (88.0%), 32x32 (96.7%) — Piezo resolution ablation. Dashed line: XL-only baseline. Higher spatial density consistently improves detection.

Zero-Shot Transfer: UMI → Tesollo

Trained on UMI linear slip, SlipSense transfers zero-shot to the Tesollo dexterous hand — over unseen articulation geometry and unseen system dynamics — achieving 100/100 detection and 95/100 closed-loop prevention.

Tesollo dexterous hand deployment with varied slip directions — Tesollo 2-finger pinch: varied objects with linear, rotational, and oblique slip — all zero-shot from UMI training.

Zero-Shot Tesollo Slip Detection

Slip Prevention

Related Work

Touch2Trace: Tactile Sensing for Dexterous Manipulation

Conference on Robot Learning (CoRL), 2026

How does the same tactile sensor contribute to downstream dexterous manipulation tasks?

Webpage

Reference

Cite SlipSense

SlipSense: Multimodal Tactile Learning for Low-Latency and Generalized Slip Detection

[1] T. Jian, A.T. Senthil Kumar, X. Li, Z. Chen, T. Dai, A. Sengul, M. Grimaldi, W. Lu, S. Nabi, and T. Yu · CoRL 2026 · arXiv:2609.15910

View arXiv

@inproceedings{slipsense2026,
  title     = {SlipSense: Multimodal Tactile Learning for
               Low-Latency and Generalized Slip Detection},
  author    = {Tong Jian and Aditya Thurvas Senthil Kumar
               and Xinyi Li and Ziling Chen and Tianyu Dai
               and Ali Sengul and Matteo Grimaldi and Wenjie Lu
               and Saleh Nabi and Tao Yu},
  booktitle = {Conference on Robot Learning (CoRL)},
  year      = {2026},
  eprint    = {2609.15910},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2609.15910}
}