DEUX MuJoCo — randomized starts within 5 cm
A simulated moving-conveyor bread pick-and-place dataset for the DEUX dual-arm robot. Stored in LeRobot v2.1 format with 10 episodes, 4,744 aligned frames at 30 FPS, and 30 camera videos. Parquet is uncompressed for reader compatibility.
Collection
- Episode seeds: 7–16. One bread target per episode.
- Both hands start at independently randomized EE positions within 5 cm of the reference home pose. Reference hand orientation is retained.
- 4 successful episodes and 6 failed episodes, all retained. The failed episodes reached the controller's small-rotation alignment limit and did not attempt a grasp. Success-only training requires filtering the episodes.
- Conveyor speed: 0.06 m/s, continuously moving during each recorded episode.
- The controller uses privileged simulator object pose and velocity.
- Simulation uses contact-gated weld grasp assistance and disables robot self-collision. No physical tool was added to the robot.
Features
- RGB:
observation.images.head,observation.images.left_wrist,observation.images.right_wrist, each 360 × 640 × 3. observation.state: 30 float32 values, measured state.action: 30 float32 values, absolute targets. EE targets are FK of commanded arm joints, rather than the next measured state.- Vector order: left EE pose (7), right EE pose (7), left arm joints (7), right arm joints (7), left/right gripper (2).
- EE pose: world-frame
[x, y, z, qx, qy, qz, qw], position in meters. Arm joints are in radians. Gripper values: 0 closed, 1 open.
Alignment and provenance
Raw simulation samples were recorded every 34 ms. Conversion uses a 30 FPS grid on the common stream interval and selects only the latest sample at or before each grid point. Sample age and source gaps are limited to 40 ms. Quaternion sign continuity is corrected without changing orientation. Raw HDF5 files are not included in this converted dataset.
alignment/ contains source indices, timestamps, initialization metadata,
stage/grasp events, and per-episode summaries. Episode outcomes and source hashes
are also listed in meta/conversion.json. Source paths describe the original
collection machine and are provenance, not required runtime dependencies.
Loading
Use a LeRobot v2.1-compatible installation (tested with upstream LeRobot tag
v0.3.3, datasets 3.6.0, Torch 2.7.1 and Torchvision 0.22.1). The repository's
v2.1 revision is provided for legacy loader compatibility. Newer LeRobot
v3-only loaders require format conversion.
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset(
"2sin0/deux-mujoco-random-start-5cm-10",
revision="v2.1",
video_backend="pyav",
)
sample = dataset[0]
The repository is public and can be downloaded without authentication. validation_v21.json records successful loading of first/middle/last
samples in every episode and checks of temporal queries and boundary padding.
All video frames were decoded during export to verify counts, dimensions, FPS,
and timestamps.
- Downloads last month
- 26