A-SLIP: Acoustic Sensing for Continuous In-hand Slip Estimation

Uksang Yoo; Yuemin Mao; Jean Oh; Jeffrey Ichnowski

cs.ROApr 9, 2026

A-SLIP: Acoustic Sensing for Continuous In-hand Slip Estimation

Authors: Uksang Yoo, Yuemin Mao, Jean Oh, Jeffrey Ichnowski

Organizations: Robotics Institute, Carnegie Mellon University, USA

Abstract

Reliable in-hand manipulation requires accurate real-time estimation of slip between a gripper and a grasped object. Existing tactile sensing approaches based on vision, capacitance, or force-torque measurements face fundamental trade-offs in form factor, durability, and their ability to jointly estimate slip direction and magnitude. We present A-SLIP, a multi-channel acoustic sensing system integrated into a parallel-jaw gripper for estimating continuous slip in the grasp plane. The A-SLIP sensor consists of piezoelectric microphones positioned behind a textured silicone contact pad to capture structured contact-induced vibrations. The A-SLIP model processes synchronized multi-channel audio as log-mel spectrograms using a lightweight convolutional network, jointly predicting the presence, direction, and magnitude of slip. Across experiments with robot- and externally induced slip conditions, the fine-tuned four-microphone configuration achieves a mean absolute directional error of 14.1 degrees, outperforms baselines by up to 12 percent in detection accuracy, and reduces directional error by 32 percent. Compared with single-microphone configurations, the multi-channel design reduces directional error by 64 percent and magnitude error by 68 percent, underscoring the importance of spatial acoustic sensing in resolving slip direction ambiguity. We further evaluate A-SLIP in closed-loop reactive control and find that it enables reliable, low-cost, real-time estimation of in-hand slip. Project videos and additional details are available at https://a-slip.github.io.

Figures & tables

Fig. 1: Overview of A-SLIP: Piezoelectric microphones embedded behind textured silicone contact pads capture structure-borne vibrations during slip. Multi-channel log-mel spectrograms are processed by a convolutional network with channel and temporal attention to jointly estimate slip presence, magnitude, and direction as $\mathbf{v}_{t}\in\mathbb{R}^{2}$ .

Fig. 2: Design of the A-SLIP Sensor. ( Left ) Fabrication pipeline: silicone is cast in a 3D-printed mold to form a compliant contact pad, which is bonded to a rigid holder with embedded piezoelectric microphones. ( Center ) Contact surface variants: smooth and textured silicone pads; the textured surface introduces controlled asperities that produce more directionally informative vibrations during slip. ( Right ) Evaluated microphone placements: three two-microphone layouts (centered, corners, and same-finger) and a four-microphone layout with two microphones per finger to increase contact-region coverage.

Fig. 3: A-SLIP Model Architecture. Log-mel spectrograms from synchronized microphones are normalized and fused via a learned channel attention module. The fused representation passes through 2D convolutional layers preserving temporal resolution, then 1D temporal convolutions modeling slip dynamics. A temporal attention pooling module aggregates features into a latent vector passed to three heads: a slip classification head predicting $p(\text{slip})\in[0,1]$ , a magnitude regression head predicting $\|\mathbf{v}\|$ , and a direction head predicting $\hat{\mathbf{v}}\in\mathcal{S}^{1}$ .

Fig. 4: A-SLIP System. We mount A-SLIP sensors on an XArm gripper attached to an XArm7 robot. To obtain ground-truth in-hand slip, we use an OptiTrack Trio to track poses of the left finger and the object, each with reflective markers attached.

Fig. 5: A-SLIP Dataset Distribution. ( Top ) Robot-induced slip data collected automatically via randomized robot motions sweeping the gripper across a stationary probe, with labels derived from robot states. ( Bottom ) Externally-induced slip data collected by manually perturbing grasped objects, with labels from OptiTrack-tracked rigid-body motion. Counts indicate the number of audio slices in the dataset.

Method	Det.	Dir. MAE	Mag. RMSE
Method	(%)	(deg)	(mm)
SVM	73.0	$19.2\pm 28.6$	$1.9\pm 2.4$
Single mic (no pretrain)	43.3	$28.5\pm 26.5$	$2.0\pm 1.0$
Single mic (pretrain + finetune)	70.2	$20.7\pm 22.4$	$2.7\pm 0.9$
Pretrain only (2-mic, centered)	60.4	$29.3\pm 21.5$	$1.5\pm 1.2$
Pretrain only (2-mic, corners)	63.1	$26.8\pm 21.1$	$1.5\pm 1.2$

TABLE I: Slip prediction model comparison. Binary slip accuracy (Det.). Dir. MAE: Slip direction MAE (Dir. MAE). Slip magnitude RMSE (Mag. RMSE).

Object	Train	Det.	Dir. MAE	Mag. RMSE
Object	Train	(%)	(deg)	(mm)
PLA Rod	Per-obj.	94.5	$26.07\pm 9.8$	$\mathbf{2.44\pm 1.20}$
PLA Rod	All-obj.	94.7	$\mathbf{25.58\pm 9.1}$	$2.72\pm 1.30$
Glass Cleaner	Per-obj.	81.1	$15.17\pm 7.4$	$0.60\pm 0.72$
Glass Cleaner	All-obj.	78.7	$\mathbf{14.38\pm 6.8}$	$\mathbf{0.58\pm 0.68}$
Chips Can	Per-obj.	71.8	$9.01\pm 6.1$	$\mathbf{0.40\pm 0.63}$

TABLE II: Per-object slip prediction results. Binary slip accuracy (Det.). Slip direction MAE (Dir. MAE). Slip magnitude RMSE (Mag. RMSE).

Fig. 6: Qualitative Evaluation of A-SLIP. Each row shows a different object; each column shows a sample evaluation frame with predicted (gray) and ground-truth (green) slip vectors overlaid on the contact image alongside per-channel log-mel spectrograms. A-SLIP accurately estimates slip direction and magnitude across objects with varying geometry and surface material, even under impulsive externally induced slip.

	Task 1: Slip-Stop		Task 2: Slip-Track
	Succ. , $\mathbf{\Delta x}$ (mm)		Pose RMSE (mm)
Object	SVM	A-SLIP	SVM	A-SLIP
PLA Rod	7/10, $39.7\pm 38.7$	10/10 , $\mathbf{16.7\pm 8.9}$	$32.4\pm 20.2$	$\mathbf{16.9\pm 12.4}$
Glass Cleaner	6/10, $13.4\pm 18.7$	10/10 , $\mathbf{10.5\pm 5.7}$	$32.4\pm 18.1$	$\mathbf{15.0\pm 9.5}$
Chips Can	8/10, $\mathbf{16.7\pm 16.4}$	10/10 , $19.8\pm 6.6$	$33.8\pm 21.7$	$\mathbf{12.9\pm 9.4}$
Mustard Container	5/10, $17.4\pm 31.3$	10/10 , $\mathbf{9.7\pm 6.7}$	$25.0\pm 14.4$	$\mathbf{20.5\pm 14.4}$

TABLE III: Task performances. Success rate (Succ.), stopping error (

\Delta x

), and RMSE of gripper-object pose along trajectory (Pose RMSE).

Fig. 7: Reactive Control. A-SLIP predicts slip direction and magnitude in real time, enabling rapid robot responses to in-hand slip. ( Left ) Task 1: the robot pushes an object against a wall and stops automatically upon in-hand slip detection. ( Right ) Task 2: as an experimenter induces slip, the robot follows the model-predicted slip vector to maintain a stable grasp.

Tactile Sensing

Tactile sensing measures contact forces, deformation, texture, and motion to help robots perceive and act through physical interaction.

See all 138 papers

SlipSense: Multimodal Tactile Learning for Low-Latency and Generalized Slip Detection
Sep 14, 2026Tong Jian, Aditya Thurvas Senthil Kumar, Xinyi Li +7Tactile Sensing Tactile Sensing for Robotic Manipulation
Dexterous AI Group, Analog Devices, Inc.
Acoustic Sensing for Universal Jamming Grippers
Feb 27, 2026Lion Weber, Theodor Wienert, Martin Splettstößer +2Robotic Tactile Sensing Tactile Sensing for Robotic Manipulation
Robotics and Biology Laboratory, Technische Universität Berlin · Science of Intelligence, Research Cluster of Excellence, Berlin · Robotics Institute Germany
Contact Status Recognition and Slip Detection with a Bio-inspired Tactile Hand
Mar 19, 2026Chengxiao He, Wenhui Yang, Hongliang Zhao +3Tactile Sensing for Robotic Manipulation Tactile Sensing
School of Mechanical Engineering, Southeast University, Nanjing 211189, China · School of Electronic Science & Engineering, Southeast University, Nanjing 211189, China

Sep 14, 2026J/K move · Enter open · S save

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

Tong Jian, Aditya Thurvas Senthil Kumar, Xinyi Li, Ziling Chen+6

Dexterous AI Group, Analog Devices, Inc.

Tactile Sensing Tactile Sensing for Robotic Manipulation Robotic Tactile Sensing Multimodal Sensor Fusion

Open paper AskPDF arXiv

Slip detection is fundamental to dexterous manipulation, yet existing systems often lack precise characterization of detection latency and cross-platform generalization. We present SlipSense, a multimodal tactile slip-detection framework built on TacV5, a compact sensor integrating a

32 \times 32

piezoresistive array operating at 240 Hz and a 3-axis MEMS accelerometer operating at 8 kHz. The piezoresistive array captures spatial pressure distributions, while the accelerometer captures friction-induced vibrations, 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 dataset of 1.4 million frames spanning 37 objects demonstrate the complementarity of the two modalities. SlipSense achieves 96.7% Macro F1 with a false-positive rate below 1.6%, detecting 76% of slip events within 23.1 ms. When trained solely on UMI data, SlipSense generalizes zero-shot to a Tesollo dexterous hand, transferring across unseen objects, distinct sensor units, and robotic platforms without retraining.

Abstract

Tactile Sensing

Explore similar work

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

Acoustic Sensing for Universal Jamming Grippers

Contact Status Recognition and Slip Detection with a Bio-inspired Tactile Hand