Jay-Ye/Z-Sight-data

Z-Sight G1 real-robot data

Teleoperated demonstrations on a real Unitree G1 humanoid, ten household tasks, recorded at 10 Hz with one egocentric camera. 557 episodes / 243,600 frames, one LeRobot v3 dataset per task (folder = task). This is the complete self-collected snapshot of 2026-10-13: it is a superset of zhaoyiqi/OmniEmbodied_Data (rev 3533636, the base episodes plus the *_straight sessions merged into their tasks) extended with the per-task split of DONG3344/4in1 (rev f98d605c, continuously recorded chair → trash → book → fruit chains cut into single-task episodes, 16 per task).

Tasks

folder	instruction	episodes	frames	episode length	camera	composition
bottle	Sort the white bottles into the right bin and the red ones into the left bin.	49	35,851	48–104 s	480×640	base
dog	Place the dog into the box.	60	9,569	10–26 s	480×640	base
plate	Stack the plates.	58	14,495	16–54 s	480×640	base
pour	Pour the material from white bottle to the glass.	46	13,822	18–46 s	480×640	base
puffs	Put the puffs box into the drawer.	49	13,566	14–49 s	480×640	base
walking	deliver the bottle on the drawer to the table	49	17,429	22–53 s	480×640	base
chair	Push the chair under the table.	70	24,806	23–58 s	720×1280	46 base + 8 straight + 16 continue
book	Put the book on the shelf.	61	33,391	2–101 s	720×1280	34 base + 11 straight + 16 continue
trash	Throw the trash into the trash bin.	58	32,951	42–95 s	720×1280	34 base + 8 straight + 16 continue
fruit	Put the fruits into the fridge.	57	47,720	57–139 s	720×1280	30 base + 11 straight + 16 continue
total		557	243,600

Every episode of a task carries the same instruction (meta/tasks.parquet, one row). The six 480×640 tasks were recorded with the G1's egocentric RealSense-type camera; the four 720×1280 tasks with a wide-FOV ZED camera.

Format

Each task folder is a standalone LeRobot v3.0 dataset (robot_type: g1, fps: 10):

column	dtype / shape	notes
observation.images.egocentric	video, H×W×3	AV1, yuv420p, 10 fps, no audio (videos/observation.images.egocentric/chunk-000/file-*.mp4)
states	float32[43]	proprioceptive state vector (state_0 … state_42)
action	float32[78]	commanded action vector (action_0 … action_77)
timestamp, frame_index, episode_index, index, task_index	scalars	standard LeRobot v3 bookkeeping

meta/ holds info.json, episodes/chunk-000/file-000.parquet (per-episode offsets and lengths), tasks.parquet, stats.json (per-column statistics) and a GR00T-style modality.json (ego_view → the camera, state = states[0:43], action = action[0:78], human.task_description → task_index). The four merged tasks also keep a merge_manifest.json recording which source files, episode offsets and parquet SHA-256 went into the merge. The per-dimension meaning of the state and action vectors follows the collection setup documented in zhaoyiqi/OmniEmbodied_Data.

Loading

from huggingface_hub import snapshot_download
root = snapshot_download("Jay-Ye/Z-Sight-data", repo_type="dataset", allow_patterns=["chair/*", "chair/**"])
# <root>/chair is a LeRobot v3 dataset: data/chunk-000/*.parquet + videos/.../*.mp4 + meta/

The videos are AV1; use a decoder with AV1 support (PyAV with libdav1d, or torchcodec).

History

snapshot	date	change
v1	2026-09	ten tasks, 455 episodes
v2	2026-10-06	*_straight sessions merged into book, chair, fruit, trash (490 episodes)
v3 (this repo)	2026-10-13	*_continue episodes split from the continuous 4-task recordings and appended (557 episodes)

The six September tasks are byte-identical across snapshots. book contains one 24-frame episode. The _continue episodes were cut from the continuous recordings on keyframes with a lossless AV1 packet remux (episode i of a chain = task i mod 4).