drisya-ego/ego-manip-v0.1

Drisya Egocentric Manipulation v0.1

Head-mounted, first-person video of people doing everyday bimanual tabletop tasks.
Every frame has 3D hand landmarks for both hands and the camera's IMU. Packaged as a ready-to-train LeRobot dataset.

Quickstart · Tasks · Features · Need more data? · Visualize

This is a sample. It comes from a private pool of 7,684 clips (41.4 h), and the pool is growing. For the full pool, more hours of these tasks or a custom collection for your own tasks, email info@drisya.ai.

Why this dataset

Vision-language-action (VLA) policies get better when you co-train them on large amounts of egocentric human video. That video needs consistent hand tracking and a robot-learning-friendly format. This release gives you exactly that:

• 🎥 First-person view. A GoPro mounted on the operator's head sees roughly what a robot's head camera would see.
• ✋ Both hands, every frame. 21 MediaPipe keypoints per hand, in both world (metres) and image coordinates, with confidence.
• 🧭 IMU. Accelerometer and gyroscope from the camera, resampled to the video frame rate (head-motion signal).
• 🤖 LeRobot v2.1 format. Loads in one line, has the same layout as robot datasets, and ships with stats for normalisation.
• 🧪 Quality checked. Every clip is checked for dropped frames, A/V sync, IMU coverage and hand visibility (see QA).

Quickstart

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("drisya-ego/ego-manip-v0.1")
print(len(ds), "frames,", ds.num_episodes, "episodes")

frame = ds[0]
frame["observation.images.ego"]        # (3, 360, 640) float tensor, RGB
frame["observation.hands.left"]        # (21, 3) world landmarks in metres (NaN if not detected)
frame["observation.hands.right_img"]   # (21, 3) normalised image coords (x, y, relative depth)
frame["task"]                          # e.g. "Fold the small blue towel in half twice"

You can also browse episodes in your browser with the LeRobot dataset visualizer.

Tasks

Towel folding · 150 episodes · 6 variants · 37 min Fold a towel in half twice. Small and medium towels in several colours.	T-shirt folding · 150 episodes · 3 variants · 71 min Fold a flat-laid T-shirt. Blue, cream and grey shirts.
Envelope insertion · 150 episodes · 9 variants · 42 min Put a letter into an envelope. Small, medium and large envelopes in many colours.	Nut & bolt · 150 episodes · 1 variants · 93 min Screw a nut onto a bolt, a fine-grained bimanual task.

Every variant at a glance

Mid-episode frame from each variant. Pink = left hand, cyan = right hand (the wearer's own hands).

Episodes per variant (click to expand)

task	variant	episodes	minutes
Envelope insertion	large white	12	3.6
Envelope insertion	large yellow	51	14.5
Envelope insertion	medium green	5	1.1
Envelope insertion	medium grey	2	0.5
Envelope insertion	medium light pink	5	1.6
Envelope insertion	medium red	27	5.3
Envelope insertion	medium white	5	1
Envelope insertion	medium yellow	17	4.5
Envelope insertion	small traditional	26	9.6
Nut & bolt	medium	150	92.6
Towel folding	medium blue	4	1.6
Towel folding	medium cream	2	0.6
Towel folding	medium gray	3	1.2
Towel folding	small blue	49	11.1
Towel folding	small red	56	13.6
Towel folding	small teal	36	9.2
T-shirt folding	blue	62	28.1
T-shirt folding	cream	41	20.1
T-shirt folding	grey	47	22.5

The 600 public episodes are drawn proportionally from every variant and spread across 43 recording days. Each episode has a natural-language instruction in meta/tasks.jsonl.

What's in each frame

key	dtype	shape
observation.images.ego	video	(360, 640, 3)
observation.hands.left	float32	(21, 3)
observation.hands.right	float32	(21, 3)
observation.hands.left_img	float32	(21, 3)
observation.hands.right_img	float32	(21, 3)
observation.hands.conf	float32	(2,)
observation.imu.accel	float32	(3,)
observation.imu.gyro	float32	(3,)
phase_id	int64	(1,)
timestamp	float32	(1,)
frame_index	int64	(1,)
episode_index	int64	(1,)
index	int64	(1,)
task_index	int64	(1,)

• observation.images.ego: head-mounted RGB, downscaled to 640×360 (H.264, 30 fps). 60 fps sources are decimated to 30 fps and 4K sources are downscaled.
• observation.hands.{left,right}: MediaPipe Hand Landmarker (Apache-2.0), 21 keypoints in world coordinates (metres) relative to the hand's centre.
• observation.hands.{left,right}_img: the same keypoints in normalised image coordinates.
• observation.hands.conf: per-hand handedness confidence [left, right]. A hand that isn't detected is NaN with confidence 0. Left and right are from the wearer's point of view.
• observation.imu.accel / gyro: GoPro GPMF accelerometer (m/s²) and gyroscope (rad/s), resampled onto frame times.
• phase_id: coarse per-task phase index (−1 = unlabelled).
• There is no action key. This is human data. Next-step hand landmarks work well as an action proxy (see the smoke test below).

Layout on disk

meta/      info.json · episodes.jsonl · tasks.jsonl · episodes_stats.jsonl
data/      chunk-000/episode_000000.parquet …
videos/    chunk-000/observation.images.ego/episode_000000.mp4 …
assets/    preview GIFs for this card (not part of the dataset)

Episode metadata (meta/episodes.jsonl) includes task_slug, variant, record_day, operator (pseudonymous), imu_usable and source_sha256 for provenance.

Collection setup

• Camera: GoPro, head-mounted, 1080p and up, 30 or 60 fps. The original footage is kept privately and was never modified.
• Environment: one controlled indoor setting with a table. Objects (towels, shirts, envelopes, letters, nuts and bolts) vary in size and colour inside each task.
• Operators: several operators. Operator IDs were not recorded, so the recording day is used as a proxy.
• Consent: every operator gave consent to be recorded and for the footage to be released. The camera is head-mounted and points at the table, so the operator's face is not normally in frame. The video was not blurred or otherwise redacted.

Quality

check	result
clips processed (full pool)	7,684 (41.4 h)
clips with usable IMU	100.0%
frames with ≥1 hand / both hands	99.8% / 93.7%
clips flagged for review	1
dropped frames, A/V desync, frame-count mismatch	checked per clip

The full QA report includes duration histograms, hand-visibility plots and overlay samples for every task.

Training smoke test: a small ResNet-18 + MLP was trained on this subset on one GPU to predict next-step hand landmarks from the current frame. It confirms the dataset loads and trains without errors. It is not a benchmark, and we make no performance claims.

Intended use

• Co-training on egocentric human data alongside robot data for VLA policies, using formats like Ego-Pi and EgoScale.
• Hand-pose forecasting, phase segmentation and representation learning.
• Not intended for identifying people or for surveillance.

Limitations

• One environment and a small set of objects, so models may not generalise to new scenes.
• Monocular RGB only: no depth and no calibrated camera intrinsics.
• Hand landmarks are model estimates (MediaPipe), not motion capture. Expect dropouts under occlusion and motion blur.
• There are no robot actions or proprioception. success is mostly unlabelled in v0.1.

Need more data?

This public release is a 600-episode sample. Drisya collects egocentric manipulation data at scale:

	Public sample (this repo)	Drisya data pool
Size	600 episodes · 4.0 h	7,684 clips (41.4 h), growing
License	CC BY-NC 4.0 (non-commercial)	Commercial licence available
Tasks	4 tasks · 19 variants	Same tasks + custom tasks on request
Format	LeRobot v2.1	LeRobot, raw video, or your format

📩 info@drisya.ai: tell us your use case, the tasks and the hours you need, and we'll get back to you.

License

Released under CC BY-NC 4.0 for non-commercial use. For commercial licensing of this sample or the full pool, contact info@drisya.ai. All processing tools are permissively licensed (Apache-2.0, MIT or BSD).

Citation

@misc{drisya_ego_manip_2026,
  title        = {Drisya Egocentric Manipulation v0.1},
  author       = {{Drisya AI}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/drisya-ego/ego-manip-v0.1}},
}