EgoForce

A large-scale egocentric human manipulation dataset with synchronized fingertip force.

EgoForce is the human tactile dataset introduced in FACT: A Multimodal Force-Action-Track Foundation Model for Human-to-Robot Dexterity. Each recording pairs a head-mounted stereo video and two wrist videos with glove measurements of per-finger force, pressure, bend and orientation for both hands, aligned frame by frame to the head video.

Authors

Zhi Cao^1,2*, Yining Hong^2*, Jiaolong Yang³, Yejin Choi², Li Fei-Fei², Jiajun Wu^2†, Huang Huang^2†

¹University of Michigan   ²Stanford University   ³Microsoft Research

^*Equal contribution   ^†Equal advising

Paper · Hand keypoint algorithm

At a glance

Recordings	18,373
Recordings with tactile data	18,311
Duration	739 hours
Tasks	25
Collectors	191
Collection dates	2026-09-02 to 2026-09-08
Streams	head stereo video, left and right wrist video, bimanual glove (force, pressure, bend, IMU)

Layout

<date>/<task>/<collector>/recording_NNN/
    head.mp4           head camera, side-by-side stereo, 1920x600, ~30 fps
    wrist_left.mp4     left wrist camera, 960x540, ~30 fps
    wrist_right.mp4    right wrist camera, 960x540, ~30 fps
    timestamps.json    per-frame times of each video
    tactile.npz        glove data aligned to head.mp4 frames
handpose_algorithm/    offline 2D/3D hand keypoint estimation for head.mp4

<collector> is a collector/session code. 62 recordings have no tactile.npz (no glove stream, or no usable glove data to screen).

Tasks

task	recordings	task	recordings
Assembly_of_Disassem_bly_and_Wooden_Shapes	1,054	Organize_and_categorize_hair_ties	1,199
Check_the_ruler	906	Origami	709
Clean_the_glasses	823	Pour_water	263
Disassemble_andAssemble_Toys	852	Quality_inspection_of_tableware	667
Disassembly_andAssembly_of_Computer_Backplane	623	Removing_andInstalling_Nuts	689
Disassembly_andAssembly_of_Shelves	395	Small_Bead_Classification	732
Fold_clothes	641	Sorting_Brushes	786
Jigsaw_Puzzle	554	Starting_andHammering_Nails	68
Number_Blocks_Set	973	Store_tableware	683
Organize_Toolbox	1,376	Store_the_beads	760
Unpacking_andAttaching_BookCovers	1,194	Wooden_Shape_Matching	199
building_blocks	944	lace_up_shoes	101
untie_the_shoelaces	1,182

Task folder names are kept exactly as recorded.

Video

• head.mp4 is side-by-side stereo: the left eye is x in [0, 960), the right eye is x in [960, 1920), each 960x600 (downscaled from 1920x1200 native).
• timestamps.json holds {"head": [...], "wrist_left": [...], "wrist_right": [...]}, the time in seconds of every frame of each video. All streams share one time origin, so frames of different videos can be matched by time.

Tactile (tactile.npz)

One row per frame of head.mp4 (N = number of head frames). L is the wearer's left hand and R the right hand.

key	dtype, shape	meaning
frame_t	float32 (N,)	frame time in seconds, same as timestamps.json["head"]
L_force, R_force	float16 (N,5,12)	finger force in newtons; 5 fingers x 12 taxels
L_press, R_press	float16 (N,5,12)	finger pressure, sensor units
L_bend, R_bend	float16 (N,5)	finger bend, one value per finger, sensor units
L_quat, R_quat	float16 (N,4)	glove IMU orientation quaternion
L_valid, R_valid	bool (N,)	a glove sample exists within tol_s of this frame
L_flag, R_flag	bool (N,5)	this finger was flagged as unreliable at this frame
tol_s	float32	alignment tolerance (0.03 s)
origin_ns	int64	absolute time origin of frame_t, nanoseconds

Processing:

• Alignment. Each head frame takes the nearest glove sample (glove ~200 Hz) within tol_s. Frames with no sample have *_valid == False and all values 0.
• Noise screening. A whole-hand CNN noise detector flags spans of sustained tactile noise per finger (score cutoff 0.9, minimum run 0.5 s). Flagged spans are marked in *_flag, and *_force / *_press of that finger are set to 0 there. Use *_flag to tell a screened zero from a true zero contact force.
• Palm. Palm sensors are not included.

import json, numpy as np

rec = "20260902/Pour_water/BNN_2828_E_01/recording_001"
z = np.load(f"{rec}/tactile.npz")
t = json.load(open(f"{rec}/timestamps.json"))["head"]

i = 300                                         # head.mp4 frame index
if z["R_valid"][i]:
    total_right = z["R_force"][i].astype(np.float32).sum()  # newtons over all fingers
    unreliable = z["R_flag"][i]                 # per finger

Download a subset with huggingface_hub:

from huggingface_hub import snapshot_download
snapshot_download("zhicao/EgoForce", repo_type="dataset",
                  allow_patterns=["20260902/Pour_water/*"], local_dir="EgoForce")

Hand keypoints

Hand keypoints are not included as data. handpose_algorithm/ contains the offline algorithm that estimates 2D/3D keypoints of both hands from head.mp4, frame-aligned with tactile.npz. See its README for setup (MANO must be downloaded separately).

License

The dataset is released under Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0): free for non-commercial use with attribution. The handpose_algorithm/ folder contains third-party code and weights under their own licenses (see handpose_algorithm/licenses/), including WiLoR (CC BY-NC-ND 4.0).

Citation

If you use EgoForce, please cite:

@article{cao2026fact,
  title   = {FACT: A Multimodal Force-Action-Track Foundation Model for Human-to-Robot Dexterity},
  author  = {Cao, Zhi and Hong, Yining and Yang, Jiaolong and Choi, Yejin and Fei-Fei, Li and Wu, Jiajun and Huang, Huang},
  journal = {arXiv preprint arXiv:2610.19155},
  year    = {2026}
}

Acknowledgements

The data were collected with the FusionX multimodal data collection system (TouchTronix Robotics, NY, USA).