BrickCraft

Single-arm Kinova Gen3 demonstrations for task-conditioned interlocking brick assembly, released in LeRobot dataset format v3.0.

• Project website
• Paper
• Model checkpoints

Dataset overview

Property	Value
Robot	Kinova Gen3, single arm
Episodes	865
Frames	52,072
Sampling rate	10 Hz
Split	train, episodes [0, 865)
Image streams	4 RGB video streams, each 256 × 256
Observation state	13D: end-effector pose (9D), gripper position (1D), scaled external force (3D)
Action	10D: target end-effector pose (9D), gripper command (1D)
Task conditioning	6D observation.task_encoding

Camera keys

The dataset contains two physical camera views and two derived fixed-camera streams. All four keys are stored as videos under videos/<camera_key>/; video frames have RGB channel order. LeRobot returns decoded image tensors in channel-first order (3 × 256 × 256).

Full camera key	Meaning
observation.images.wrist_camera	Cropped and resized RGB view from the wrist-mounted camera.
observation.images.fixed_camera_raw	Raw visual observation from the fixed camera.
observation.images.fixed_camera	Mask overlay with background dimming.
observation.images.fixed_mask	Semantic segmentation mask from the fixed camera.

fixed_camera, fixed_camera_raw, and fixed_mask share the same cropped fixed-camera view and frame alignment. They are alternative representations of that view, rather than three separate cameras. Select the image keys expected by your policy checkpoint.

State, action, and force scaling

observation.state is a float32 vector with 13 entries in this order:

[ee_pose_x, ee_pose_y, ee_pose_z,
 ee_pose_r11, ee_pose_r12, ee_pose_r13,
 ee_pose_r21, ee_pose_r22, ee_pose_r23,
 gripper_position,
 fx_scaled, fy_scaled, fz_scaled]

The first nine entries represent the end-effector pose. Position is in meters. Rotation uses the column-wise 6D representation: the first three rotation entries are R[:, 0] and the next three are R[:, 1]. The stored feature names above are preserved from the dataset; the two rotation triples are the first two columns of the rotation matrix. Observation pose comes from forward-kinematics feedback.

External force. The final three state entries store external force in newtons multiplied by 0.05; multiply the stored values by 20 to recover newtons.

action is a float32 vector with 10 entries: the same 9D pose representation followed by the gripper command. Pose actions are Cartesian targets rather than pose deltas or joint angles. Gripper commands use 0.0 for open and 0.8 for close; the observed gripper position is feedback and can take intermediate values.

Task encoding

Refer to the paper for an explanation of observation.task_encoding.

Download and load

With the Hugging Face CLI installed:

hf download JerryYuu/brickcraft --repo-type dataset --local-dir ./dataset/brickcraft

With LeRobot 0.4.4 and a compatible video decoder installed:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset(
    repo_id="JerryYuu/brickcraft",
    root="./dataset/brickcraft",
)
sample = dataset[0]
print(sample["observation.state"].shape)         # torch.Size([13])
print(sample["action"].shape)                    # torch.Size([10])
print(sample["observation.task_encoding"].shape) # torch.Size([6])
force_newtons = sample["observation.state"][10:13] * 20.0

Use a decoder that supports both AV1 and H.264. For reproducibility, pin downloads to the dataset repository commit used in your experiment.

Files

data/chunk-000/file-000.parquet
meta/info.json
meta/stats.json
meta/tasks.parquet
meta/episodes/chunk-000/file-000.parquet
videos/observation.images.wrist_camera/chunk-000/file-000.mp4
videos/observation.images.fixed_camera/chunk-000/file-000.mp4
videos/observation.images.fixed_camera_raw/chunk-000/file-000.mp4
videos/observation.images.fixed_mask/chunk-000/file-000.mp4

The Parquet files contain frame data, episode ranges, timestamps, video offsets, task mapping, and feature statistics. Each video file contains multiple episodes; use episode metadata to locate the correct timestamps. SHA256SUMS records the release file checksums.

Scope

This dataset contains single-arm assembly demonstrations for the seven task encodings listed above. It does not include a held-out validation or test split. Segmentation masks are derived annotations and may contain tracking errors. Training and evaluation should use compatible state dimensions, task encodings, image representations, and force scaling.

License

This dataset is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt the data, including for commercial purposes, subject to the license terms: give appropriate credit, provide a link to the license, and indicate if changes were made. See LICENSE for the license reference.

When redistributing the dataset or adapted data, attribute the BrickCraft dataset and link to https://huggingface.co/datasets/JerryYuu/brickcraft. This dataset license applies to the data distributed here; the separately released software and model checkpoints have their own licenses.