jeeekgao/a3-cookie-2x10-advanced-4000

A3 Cookie Transfer — 2×10, Advanced, 4000 Episodes

4000 successful simulated expert demonstrations of packing 20 upright cookies into a two-column, ten-row box with an A3 dual-arm robot. The left arm performs the task; the right arm remains at its deployment home. This is a simulation dataset, with three synchronized RGB views and joint position actions in LeRobot v3.0 format.

Item	Value
Episodes / frames / language tasks	4000 / 4,464,136 / 40
Control frequency	20 Hz
Duration	62.00 hours per camera
Views	front, left wrist, right wrist; each 256×256 RGB
State / action	16D / 16D
Video	AV1 MP4, yuv420p, no audio
Randomization	advanced; episode-stable
Recorded expert	cookie_2x10_variable_grasp
Collector code	3b0b06d406d35c223da00a850efa32261d6c3838
Verified environment	LeRobot 0.5.1, MuJoCo 3.14.0

Task and instruction

There are 80 cookies in four source columns, 20 per column. A source column in 1–4 and a first-grasp count n in 0–9 determine the instruction. The four planned quantities are [n, 10-n, n, 10-n]; the first two fill target column 1 and the next two fill target column 2. Zero quantities skip that grasp. random samples n once per episode; the first and third quantities match.

Example (source_column=1, n=3):

Transfer 20 cookies from source column 1 into the 2x10 target box. Grasp 3 cookies first, then 7 to fill target column 1 with 10. Repeat: grasp 3, then 7 to fill target column 2 with 10. Release all 20 cookies upright in the target box and leave 60 in the source box.

Successful demonstrations require 20 distinct confirmed transfers, release, upright stable placement, target counts [10,10], and 60 cookies left in the source. Failed collection attempts are excluded from the indexed learning episodes. Shared video files can contain unindexed intervals retained by the recorder; LeRobot uses episode timestamps and ignores these intervals. Their durations are documented in metadata/video_audit.json. Expert controls use privileged simulated object positions and real contact dynamics; policy observation fields do not include that privileged metadata.

Distribution

Each source column has exactly 1000 successful episodes. First-grasp counts cover every value 0–9 but are not exactly balanced after success filtering.

Source column	n=0	n=1	n=2	n=3	n=4	n=5	n=6	n=7	n=8	n=9	Total
1	111	110	82	94	108	99	88	106	113	89	1000
2	93	111	96	81	81	97	117	106	108	110	1000
3	107	111	108	116	76	83	104	85	109	101	1000
4	117	97	98	112	109	103	103	87	83	91	1000

Colors: berry: 970, cocoa: 1000, golden: 988, matcha: 1042. Full episode index, seed, quantity, color, duration and task index: episodes.csv.

Observation and action

observation.state and action order: [L1..L7, L_gripper, R1..R7, R_gripper]. For each arm the joint order is shoulder pitch, shoulder roll, shoulder yaw, elbow pitch, wrist roll, wrist pitch, wrist yaw. Joints are in radians; gripper commands use 0=closed and 1=open. Simulated measured gripper state may slightly overshoot this nominal range.

Actions are absolute joint position targets, recorded after IK and safety/rate processing, paired with the observation before the action. Do not interpret them as Cartesian deltas, joint increments, velocity or torque.

Feature	Shape	Meaning
observation.images.front	256×256×3	RGB external camera
observation.images.left_wrist	256×256×3	RGB left wrist camera
observation.images.right_wrist	256×256×3	RGB right wrist camera
observation.state	16	Measured joint positions and gripper openings
action	16	Applied absolute targets and gripper commands
observation.velocity	16	State rates
observation.eef_pose	14	Left XYZ + WXYZ quaternion, then right XYZ + WXYZ; world-frame positions in metres
observation.force	18	Left force XYZ/torque XYZ, right force XYZ/torque XYZ, four finger touch signals, two summed finger actuator forces
timestamp, frame_index, episode_index, index, task_index	scalar	Time and relational indices

The complete schema and joint names are in schema.json. meta/stats.json contains the original normalization statistics. The right-arm state and action are constant across all demonstrations. Deployment should hold that demonstrated pose; tiny deviations in zero-variance channels can be strongly amplified by normalization. This dataset does not teach active right-arm behavior.

Camera and randomization

Nominal front camera position: (0.50, -0.18, 1.25) metres; look-at point: (0.15, 0.23, 0.77) metres; vertical field of view: 48 degrees. This is the original elevated oblique view used for collection. Wrist field of view: 70 degrees. Images are native 256×256; no higher-resolution detail is implied by model resizing.

Advanced randomization includes ±20 mm source/target box translation, ±0.08 rad target yaw, ±0.3 mm cookie position noise, ±0.015 rad cookie yaw, ±9 mm front camera and ±3 mm wrist camera position perturbation, ±2/1 degree front/wrist rotation perturbation, ±2 degree FOV, lighting variation 0.2 and a four-color cookie palette. Camera, light and color samples remain fixed within an episode. Exact settings and per-episode realized camera/scene parameters are preserved in the metadata.

Load and train

Install LeRobot 0.5.1 (the collector/loader version used to verify this release), then load the Hub dataset:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo_id = "YOUR_USERNAME/YOUR_DATASET"
dataset = LeRobotDataset(repo_id)
sample = dataset[0]
print(dataset.num_episodes, dataset.num_frames)

For a local download, use LeRobotDataset(repo_id, root="/path/to/dataset"). For SmolVLA, use the three image keys, 16D state and 16D joint position action with the instruction from the dataset loader. Read video timestamps through LeRobot metadata: multiple episodes share a video shard, so a filename is not an episode ID. Do not train on the privileged a3_episode_metadata.jsonl geometry as policy input.

Optional train / validation / test partition

For compatibility, meta/info.json exposes all 4000 episodes as the original train split. suggested_splits.json additionally provides deterministic episode-level lists: 3600 train, 200 validation, 200 test. Each of the 40 source/count combinations contributes five validation and five test episodes. Apply these lists explicitly, for example:

import json
from pathlib import Path
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = Path("/path/to/dataset")
splits = json.loads((root / "metadata/suggested_splits.json").read_text())["episodes"]
train = LeRobotDataset(repo_id, root=root, episodes=splits["train"])
validation = LeRobotDataset(repo_id, root=root, episodes=splits["validation"])

These lists are for new training runs. The existing project 100k-step checkpoint was trained on all 4000 episodes; its results on these lists are not held-out results. To obtain a fully independent data split, compute normalization statistics on the training episodes only and preserve those statistics with the new checkpoint; this release retains the original all-episode statistics for reproducibility. Use independent simulator seeds for closed-loop task evaluation.

Verification and release contents

All 4,464,136 frame records were checked for finite vector values, expected dimensions, contiguous indices, task consistency and timestamps. All 239 video containers were checked, and 717 native frames (first/middle/last per container) were decoded. This is sampled video decoding, not exhaustive frame decoding. audit_report.json records the checks; SHA256SUMS and manifest.json cover the release files except the manifest/checksum files themselves.

Learning Parquet/MP4 files and canonical meta/ files are byte-identical to the merged dataset. Supplemental metadata replaces server-local source-shard paths with portable identifiers. Raw duplicate shards, attempt logs, checkpoints, credentials and training logs are excluded. Collection attempt counts describe expert collection, not learned-policy success.

Provenance and use

Source project: Eter0109/a3_dual_arm_sim. See provenance.json, asset source notes and the preserved Robotiq mesh notice. No dataset license has yet been assigned by the owner; the Robotiq MIT notice does not establish a license for A3 assets or the dataset. The package contains rendered data and metadata, not the robot mesh bundle.

This dataset is suitable for research on simulated cookie packing, imitation learning, multi-view control and instruction-conditioned variable-size grasps. Successful expert demonstrations do not imply a trained model will complete the full task. Validation should use full task completion with stable release, rather than training loss or instantaneous cookie counts alone.