EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video

This repository contains the official dataset for:

EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video
ICML 2026 Spotlight

• Paper: https://arxiv.org/abs/2606.09243
• Project Page: https://egotactile.github.io/

Dataset Summary

EgoTactile is a large-scale benchmark that pairs egocentric RGB video with synchronized full-hand pressure measurements during everyday object grasping.

The dataset is designed to support research on estimating dynamic grasp pressure from visual observations. This task is challenging because hand-object contact regions are frequently occluded, while visually similar grasping observations may correspond to different pressure distributions.

EgoTactile includes:

• 12 participants
• 63 everyday objects
• 7 object categories
• Egocentric RGB video recorded at 1280 × 720 resolution and 15 FPS
• Full-hand tactile pressure measurements from 162 sensing locations
• Pressure measurements within a 0–350 N range
• Object and participant metadata
• Temporally synchronized visual and tactile streams at 15 Hz
• A bare-hand subset for evaluating transfer to natural hand appearances

Associated Methods

The accompanying paper introduces two methods evaluated on EgoTactile:

• EgoPressureFormer, a discriminative baseline for full-hand grasp-pressure estimation from egocentric video.
• EgoPressureDiff, a conditional diffusion framework that adapts a pretrained video diffusion backbone for pressure estimation under partial visual observations and physical ambiguity.

Please refer to the paper for complete methodological and experimental details.

Dataset Organization

The dataset consists of two primary subsets corresponding to different acquisition protocols.

1. Gloved-Hand Set

The Gloved-Hand Set contains synchronized egocentric RGB videos and tactile pressure measurements collected while participants wear the tactile sensing glove.

Purpose

This subset provides direct supervision for learning mappings from egocentric video observations to full-hand pressure distributions.

Contents

• Egocentric RGB video frames
• Synchronized 162-dimensional pressure measurements
• Object metadata
• Anonymized participant metadata
• Temporal and sequence identifiers

2. Bare-Hand Set

The Bare-Hand Set is designed to evaluate transfer to natural hand appearances without a visible tactile glove.

During data acquisition, the hand visible to the egocentric camera is bare, while a synchronized off-camera gloved hand performs the corresponding grasping action and provides the tactile pressure reference. The two actions are coordinated using metronome guidance.

Purpose

This subset supports evaluation of transfer from instrumented gloved-hand observations to natural bare-hand scenarios.

Contents

• Egocentric RGB videos of bare-hand grasping
• Synchronized tactile pressure references
• Object metadata
• Anonymized participant metadata
• Temporal and sequence identifiers

Modalities

Egocentric RGB Video

• Resolution: 1280 × 720
• Frame rate: 15 FPS
• Viewpoint: head-mounted egocentric camera
• Content: hand-object grasping interactions

Tactile Pressure

• Number of sensing locations: 162
• Sampling rate: 15 Hz after synchronization
• Pressure range: 0–350 N
• Coverage: full-hand tactile sensing
• Alignment: temporally synchronized with the RGB video stream

Object Metadata

Object metadata includes:

• Object name
• Object category
• Weight
• Surface material
• Fill state

Participant Metadata

Participant metadata includes:

• Anonymized participant ID (p001–p012)
• Gender
• Hand length

Participant identities are not included in the released dataset.

Intended Uses

EgoTactile is intended for research in areas including:

• Grasp-pressure estimation
• Vision-based tactile inference
• Egocentric hand-object interaction understanding
• Multimodal representation learning
• Tactile sensing
• Robotic manipulation
• Human-robot interaction
• Transfer from instrumented hands to natural bare hands

Limitations

Users should consider the following limitations:

• The dataset contains a finite set of participants, objects, and grasping behaviors.
• Data were collected under controlled acquisition conditions.
• Pressure references in the Bare-Hand Set are obtained through synchronized paired actions rather than direct sensing on the visible bare hand.
• Performance on unseen environments, camera configurations, object types, and manipulation behaviors may differ from the reported benchmark results.
• Participant-level attributes should not be used for identity inference or unintended profiling.

License

EgoTactile is released under the Creative Commons Attribution-NonCommercial 4.0 International License, abbreviated as CC BY-NC 4.0.

The dataset may be used for non-commercial research purposes with appropriate attribution. Users are responsible for complying with the license terms.

Citation

Please cite the following paper when using EgoTactile:

@article{zeng2026egotactile,
  title   = {EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video},
  author  = {Zeng, Yuan and Shi, Yujia and Tan, Tiao and Li, Xingting and Qin, Yaqi and Lu, Zongqing and Yang, Wenming and Xue, Jing-Hao and Liao, Qingmin},
  journal = {arXiv preprint arXiv:2606.09243},
  year    = {2026}
}

Contact

For questions about the dataset, benchmark, or accompanying paper, please refer to the contact information provided on the project page or open an issue in the corresponding public repository.