DEUX MuJoCo — native URDF head-camera view
One freshly collected, successful moving-conveyor bread pick-and-place episode, using the original URDF head-camera orientation. LeRobot v2.1, 979 aligned frames at 30 FPS, three RGB camera videos. Parquet is uncompressed for compatibility with readers lacking ZSTD support.
Head-camera correction
observation.images.head is rendered by top_cam, mounted at
head_camera_pose. The camera uses the original URDF optical direction, with
the ROS-frame to MuJoCo-camera conversion. Its local quaternion (wxyz) is
[-0.5, -0.5, 0.5, 0.5], local position is zero, and the mount is unchanged.
The head camera is not re-aimed at the task. Wrist cameras retain their
existing task-directed orientation.
This is a new collection, not relabeled video from the earlier
2sin0/deux-mujoco-random-start-5cm-10 dataset, whose head view was re-aimed.
Collection conditions
- Seed 7; one episode, one bread target, successful collection into the left box.
- Both EE starting positions randomized within 5 cm of reference home, retaining reference orientation. Actual initial distances: approximately 2.17 cm left, 4.08 cm right.
- Conveyor moves continuously at 0.06 m/s; no belt stop during collection.
- Controller uses privileged simulator object pose and velocity.
- Simulation uses contact-gated weld grasp assistance and disabled robot self-collision. No physical tool was added.
- 960 raw control samples at 34 ms intervals. Raw HDF5 is not included here.
Features and alignment
Three RGB video features, each 360 × 640 × 3:
observation.images.head, observation.images.left_wrist,
observation.images.right_wrist.
observation.state and action are float32[30]: left EE pose (7), right EE
pose (7), left arm joints (7), right arm joints (7), left/right gripper (2).
EE pose is world-frame [x,y,z,qx,qy,qz,qw], positions in meters, joints in
radians, gripper 0 closed / 1 open. Observations are measured states. Actions
are absolute targets; EE actions are FK of the commanded joint targets.
Conversion selects only the latest source sample at or before each 30 FPS
target timestamp, without extrapolation. Maximum age/source gap is 40 ms.
Quaternion sign continuity is enforced without changing rotation.
alignment/ stores source index mappings, camera metadata, initialization,
events and task results. Source filesystem paths are provenance only.
Loading and validation
The repository is public. Use a LeRobot v2.1-compatible installation, tested
with upstream tag v0.3.3, datasets 3.6.0, Torch 2.7.1 and Torchvision 0.22.1.
Newer v3-only loaders require format conversion.
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset(
"2sin0/deux-mujoco-urdf-head-1",
revision="v2.1",
video_backend="pyav",
)
sample = dataset[0]
validation_v21.json records official-loader checks of first/middle/last
frames, all camera features, state/action values, temporal queries and
boundary padding. Export also decoded every video frame to check counts,
dimensions, FPS and timestamps. Native head orientation and uncompressed
Parquet were independently checked before upload.
- Downloads last month
- 31