xPhi-Ego-100 is free to use under the Apache-2.0 license. Please share a few details so XPhi Robotics can understand who is using the dataset and contact you about updates, extended data and custom collection.

xPhi-Ego-100

xPhi-Ego-100 is an egocentric (first-person) video dataset of real human manual work, built for training robot foundation models and physical AI. It has 100.1 hours of head-mounted video from 152 skilled workers in 8 trades at 8 industrial and craft sites in India. Every clip comes with synchronized head IMU.

Dataset Statistics

Attribute	Value
Total hours	100.1
Total frames	10,815,632
Video clips	2,012
Median clip length	180 s (about 3 min)
Participants (workers)	152
Environments / sites	8 sites
Task categories	8 trades (12.5 hours each)
Resolution / FPS	1920x1080 @ 30 fps, HEVC (H.265), no audio
Camera	Head-mounted XphiCapture-1
Additional sensors	IMU: 3-axis accelerometer + 3-axis gyroscope, about 29.4 Hz
Total size	195.5 GB in 100 shards
Recorded	2025-04 to 2026-05

Tasks

The hours are split equally across tasks. Within each task, clips are spread across as many workers as possible.

`task`	Work shown	Hours
`electronics_assembly`	soldering, component placement, inspection, wiring	12.5
`gem_and_jewelry`	diamond cleaving, bruting, cutting, polishing, grading	12.5
`metal_fabrication_and_machining`	cutting, grinding, welding, bending sheets and rods	12.5
`pulp_and_paper`	pulping, sheet forming, roll handling, cutting, packing	12.5
`semiconductor_fabrication`	wafer handling, photolithography prep, etching, wire bonding, electrical testing	12.5
`stone_and_masonry`	granite cutting, chiseling, edge shaping, polishing, slab handling	12.5
`textile_weaving`	warping, threading, weaving, defect fixing, beam handling	12.5
`woodworking_and_furniture`	sawing, sanding, drilling, assembly, finishing	12.5

Why xPhi-Ego-100

Real work, not staged: recorded during genuine manual tasks by professional workers at their own workplaces.
Hands in view: the head-mounted camera looks at the work area, so hands, tools and workpieces are usually in frame.
Synchronized IMU: head motion recorded alongside every clip.
Robot-ready: clip-level metadata (task, environment, difficulty, anonymous worker and site IDs) for imitation learning and video-to-action pretraining.

Dataset Structure

Data is stored as WebDataset shards. Clips are in random order, so every shard mixes tasks.

xPhi-Ego-100/
├── README.md
├── data/
│   ├── shard-000000.tar      # ~2 GB each
│   ├── shard-000001.tar
│   └── ... shard-000099.tar
└── metadata/
    └── clips.parquet          # one row per clip, incl. the shard that holds it

Inside each .tar shard, each clip is three files that share a key:

<clip_id>.mp4            # egocentric video
<clip_id>.imu.parquet    # IMU samples
<clip_id>.json           # metadata

Example <clip_id>.json:

{
  "clip_id": "xphi_000001",
  "participant_id": "participant_095",
  "site_id": "site_08",
  "task": "metal_fabrication_and_machining",
  "task_description": "Metal fabrication work including cutting, grinding, welding, and bending metal sheets and rods.",
  "environment_l1": "Manufacturing",
  "environment_l2": "Metal Fabrication",
  "environment_l3": "Sheet-Metal Fabrication & Assembly Plant",
  "task_difficulty": "medium",
  "human_interaction": false,
  "handedness": "Right-handed",
  "country": "IN",
  "date": "2026-03-28",
  "device": "XphiCapture-1",
  "duration_s": 180.0,
  "fps": 30.0,
  "resolution": "1920x1080",
  "imu_samples": 5400,
  "imu_rate_hz": 29.4
}

Metadata fields (`metadata/clips.parquet` and each `.json`)

Field	Description
`clip_id`	Unique clip ID, e.g. `xphi_000123`
`participant_id`	Anonymous worker ID, stable across that worker's clips
`site_id`	Anonymous site ID (`site_01` … `site_08`)
`task`	One of the 8 tasks above
`task_description`	Short description of the work
`environment_l1`, `environment_l2`, `environment_l3`	Environment from general to specific
`task_difficulty`	`easy`, `medium` or `hard`, as labeled by the collection team
`human_interaction`	`true` if the clip involves working with another person
`handedness`	`Right-handed`, `Left-handed` or `Ambidextrous`
`country`	`IN`
`date`	Recording date (local time, India), `YYYY-MM-DD`
`device`	`XphiCapture-1`
`duration_s`	Clip length in seconds
`fps`	Video frame rate
`resolution`	`1920x1080`
`imu_samples`	Number of IMU samples in the clip
`imu_rate_hz`	Median IMU sample rate
`shard`	The shard that holds the clip (`clips.parquet` only)

IMU

Each .imu.parquet file has one row per sample:

Column	Type	Unit	Description
`t_s`	float64	s	Time since the clip's first IMU sample. Use this to align with the video.
`t_us`	int64	µs	Raw device timestamp, as recorded (the device clock, not the video clock)
`acc_x`, `acc_y`, `acc_z`	float32	m/s²	Accelerometer, raw (includes gravity)
`gyro_x`, `gyro_y`, `gyro_z`	float32	rad/s	Gyroscope, raw

Axes: the camera frame. x is right, y is down, z is forward from the camera, right-handed (the OpenCV convention). The sensor is on the head-mounted device.
Rate: about 29.4 Hz (median interval about 34 ms), roughly one sample per video frame.
Calibration: values are raw, so expect small biases (at rest, |acc| can read slightly above 9.81 m/s²).
Sync with video: the IMU and the video start together. Video frame i is at i / 30 s, and an IMU sample is at t_s.

Quick Start

Request access on this page first (the form is approved automatically), then log in with hf auth login.

Stream with datasets:

from datasets import load_dataset

ds = load_dataset("xphirobotics/xPhi-Ego-100", split="train", streaming=True)
for sample in ds.take(1):
    print(sample["json"])
    video_bytes = sample["mp4"]
    imu_bytes = sample["imu.parquet"]

Browse the metadata, then download only the shards you need:

import pandas as pd
from huggingface_hub import hf_hub_download

repo = "xphirobotics/xPhi-Ego-100"
clips = pd.read_parquet(hf_hub_download(repo, "metadata/clips.parquet", repo_type="dataset"))
weaving = clips[clips.task == "textile_weaving"]
for shard in sorted(weaving.shard.unique()):
    hf_hub_download(repo, shard, repo_type="dataset")

Stream with webdataset:

import io, json
import pandas as pd
import webdataset as wds
from huggingface_hub import get_token

url = ("https://huggingface.co/datasets/xphirobotics/xPhi-Ego-100/resolve/main/"
       "data/shard-{000000..000099}.tar")
url = f"pipe:curl -s -L -H 'Authorization:Bearer {get_token()}' {url}"

for sample in wds.WebDataset(url, shardshuffle=False):
    meta = json.loads(sample["json"])
    imu = pd.read_parquet(io.BytesIO(sample["imu.parquet"]))
    video_bytes = sample["mp4"]  # decode with PyAV, torchvision, decord, ...
    break

Data Collection

Workers wore the XphiCapture-1 head-mounted camera while doing their normal work at 8 industrial and craft sites in Bengaluru and Ahmedabad, India. Recordings were split into clips of about 3 minutes. Clips were chosen for this release from a larger pool. Clips were picked to balance hours across tasks and spread them across workers.

Composition

Share of hours (participant-level attributes, reported only in aggregate):


Gender	male 88%, female 12%
Age	25–30: 42%, 30–34: 21%, 35–39: 25%, 45–49: 12%
Handedness	right-handed 89%, ambidextrous 11%, left-handed <1%
City	Bengaluru 75%, Ahmedabad 25%
Working with another person	4% of hours
Difficulty	hard 56%, medium 34%, easy 10%

License

Released under the Apache-2.0 license. Free for research and commercial use.

Privacy

Every participant gave recorded consent to data collection. All participants are adults.
Audio was removed from all videos.
All embedded file metadata (creation times, device and encoder tags) was stripped.
Workers and sites have new anonymous IDs. No names, worker IDs, device IDs, GPS or geolocation data are included. Location is given only at country level per clip.
The videos show workers' hands, tools and workpieces from a first-person view.

Citation

@dataset{xphi_ego_100_2026,
  title     = {xPhi-Ego-100: Egocentric Video of Real Manual Work},
  author    = {XPhi Robotics},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/xphirobotics/xPhi-Ego-100}
}

Contact

Need more hours, specific tasks or custom data collection? Visit xphirobotics.com.

Downloads last month: 138

Total file size:

196 GB

Request access to xPhi-Ego-100