xingyoujun/agibot_dataset_v1

Agibot G2 manipulation dataset v1

Five separate G2 manipulation datasets in LeRobot v2.1 format. Load each task directory as its own dataset root; the repository root is a collection of tasks, not a single combined LeRobot dataset.

Directory	Task	Episodes	Frames	Camera videos
stack_bowls	Stack three bowls	200	241,647	600
peg_into_sleeve	Insert the cylindrical peg into the fixed sleeve	200	227,032	600
clean_workcell_table	Put aluminum stock in the blue bin and the drill in the yellow bin	200	320,400	600
pick_apple	Put the apple onto the plate	200	91,883	600
pick_orange	Put the orange into the bowl	200	102,532	600
Total		1,000	983,494	3,000

Data contract

• Robot type: agibot_g2_omnipicker.
• Control/video rate: 15 Hz.
• observation.state: 60 float32 values (46 measured robot joints followed by right and left end-effector xyz + xyzw quaternion).
• action: 16 float32 values (right arm joints[7], right binary gripper, left arm joints[7], left binary gripper). Joint targets are absolute; +1 opens the gripper and -1 closes it.
• Three synchronized H264 RGB streams: head 640x400; left and right wrist 640x528.
• Recorded convention: state and images at time t precede action[t]. Inspect each task's meta/g2_conventions.json for implementation-specific details.

Each task contains data/chunk-000/*.parquet, videos/chunk-000/observation.images.*/ and meta/ (task instructions, episode records, statistics and feature names).

Collection and review

The bowl and peg tasks are retained historical G2 datasets. Cleanup was collected in October 2026 with native cuMotion, states-only recording followed by three-camera replay. It contains exactly two objects and two bins at a fixed layout; different seeds do not imply randomized scenes. Drill placement has no required final yaw. A successful episode requires complete placement and settled objects with open grippers.

pick_apple and pick_orange were collected in October 2026 with the data engine (left arm only, states-only recording followed by three-camera replay). Only the object start positions are randomized; there is no yaw, lighting, material, camera or distractor randomization. Each dataset holds the first 200 successful episodes from five 50-episode batches (apple 223/250 successful, orange 246/250), with still frames removed per the data engine's rules.

Cleanup raw task checks passed for all 200 episodes. Videos are re-rendered from recorded states. Per-episode replay reports and source audit files remain local and are not included in this training-data upload.

UPLOAD_MANIFEST.json records the uploaded payload file sizes and SHA256 digests. Local Hugging Face caches, raw HDF5, simulation assets, credentials and collection scripts are excluded. The upload contains only training data, videos, basic LeRobot metadata, this README and the upload file manifest.