RoboTwin five-task end-effector dataset — LeRobot v3.0
This release contains 2,750 ALOHA-AgileX episodes and 856,023 training frames:
50 clean and 500 randomized episodes for each task below. All episodes are in
the train split; no independent validation or test split is supplied.
complete: true in meta/build_state.json records conversion scope completion.
Quality checks and their coverage are recorded separately in meta/repair_validation.json.
| Task | Manipulation coverage |
|---|---|
| adjust_bottle | Pose and orientation adjustment |
| beat_block_hammer | Tool use and contact |
| handover_block | Bimanual handover |
| open_microwave | Articulated-object constraint |
| stack_blocks_three | Sequential precise placement |
Source: TianxingChen/RoboTwin2.0, revision
3dc3b798668feb99ac61cc9086d84cbcc3d79186, only the ten selected source ZIPs.
Unified training layout
The main branch contains standard LeRobot v3 data/, videos/ and meta/
directories, with global episode/frame numbering and shared files, independent
of task folder names. Parquet/video shards are compacted with LeRobot's official
aggregation implementation (100 MB / 200 MB targets). Metadata contains the
episode boundaries and video offsets. The data represent one multi-task dataset.
task_index identifies a language instruction; source.task_id identifies the
RoboTwin scenario via meta/source_tasks.json. source.variant_id is 0 for clean,
1 for randomized. Source episode/frame IDs remain available as columns.
The research branch retains auxiliary HDF5, episode provenance, all original
language alternatives, seeds, opaque trajectory files, robot URDF/SRDF/meshes and
converter code. Use the exact commit in meta/research_files.json to retrieve
the matching version. Those files are not required for ordinary policy training.
No raw pickle is deserialized by the converter.
State, action and geometry
Primary state and action each have 16 elements:
left_x left_y left_z left_qw left_qx left_qy left_qz left_gripper
right_x right_y right_z right_qw right_qx right_qy right_qz right_gripper
Positions are metres in the original SAPIEN world frame; quaternion order is
wxyz, normalized with continuous signs. The reference is RoboTwin's recorded
end-effector frame, not its offset TCP frame. Exact coordinate conventions and
source-code revisions are in meta/coordinate_conventions.json.
State at t contains recorded end pose and commanded normalized gripper opening. Action is the recorded pose/gripper at t+1: a next-sample demonstration waypoint, not a separately logged Cartesian controller command. N raw frames yield N-1 training frames. Full-precision original values and terminal frames are retained on the research branch.
Additional Parquet features preserve the original 14-D joint drive targets,
both end-effector world transforms, and all four camera intrinsic/extrinsic
transforms. These joint values are commands, not measured qpos. Camera transforms
map OpenCV camera coordinates to world coordinates and are checked against the
source OpenGL convention. Matrices are flattened row-major and act on column
vectors. For a reference frame F, use inv(T_world_F) @ T_world_ee.
End-effector reference verification (2026-09-28)
For the versioned aloha-agilex URDF, the recorded EE pose corresponds to
the origin and orientation of fl_link6 / fr_link6 (flange/EE reference),
expressed in the original world frame. The gripper TCP is displaced by
0.12 m along the EE's local +X axis:
p_world_tcp = p_world_ee + R_world_ee @ [0.12, 0, 0].
EE and TCP have the same orientation in this convention.
The audit checked both arms in all 856,023 training frames, plus two original HDF5 episodes. Joint-command FK gives a median flange position error of 0.454 mm (95th percentile 6.225 mm). Drive targets are not measured joint angles: the maximum error is 242.835 mm, so these fields are unsuitable as exact measured-qpos ground truth. Independently recovering link6 from wrist camera extrinsics and the fixed URDF camera mount gives a maximum position discrepancy of 0.000567 mm; treating the same poses as TCP gives about 120 mm. These are numerical simulator-consistency checks, not hardware accuracy claims.
The former ee_link_orientation description confused SAPIEN's joint global
frame with the URDF link frame. The stored quaternion already matches link6;
applying global_trans_matrix again to URDF FK introduces a 180-degree error.
The convention documentation is corrected; no recorded pose is transformed.
Detailed results: meta/eef_reference_validation.json.
Images and timing
Three cameras: observation.images.cam_high, cam_left_wrist, cam_right_wrist.
Images retain the source 320x240 resolution and use H.264 (CRF 18, GOP 2). There
is no artificial upscale to the reference unified dataset's 640x480 resolution.
The fourth camera's original encoded RGB is preserved on the research branch.
Primary video encoding is lossy; pinned source archives contain the original JPEGs.
30 FPS is nominal for LeRobot indexing, not a measured physical sampling rate. Legacy source files lack timestamps; no resampling is performed.
Repair history and validation
The original 2026-09-27 release had three one-frame video omissions in open_microwave/randomized (global episodes 1713, 1893 and 1964) and inaccurate numeric standard deviations caused by float32 variance cancellation. The 2026-09-28 repair restores those complete camera sequences from pinned source HDF5 files, corrects the affected offsets, and recomputes every numeric feature's global and per-episode statistics using float64 centered variance. Numeric quantiles are exact linear quantiles. All 856,023 numeric rows are unchanged. Unaffected video content in the repaired shards is copied without re-encoding.
The old meta/compaction_validation.json and research meta/validation/ files
describe historical checks; they did not detect these defects and are retained
for provenance. Use meta/repair_validation.json for the repaired release's
checks, hashes, and explicit validation limits. Early commits and the preserved
pre-repair tag still contain the original defects.
The Hub Dataset Viewer displays only the 23 numeric frame columns selected by
the YAML data_files configuration. It is not a robot trajectory/video player;
LeRobot joins the MP4 videos using episode metadata. The 1,323 task_index
entries are distinct selected language instructions, not 1,323 robot tasks.
There are exactly five source scenarios.
Loading
from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("MikeHan517/robotwin-eef-lerobot", revision="main")
Use a fresh local dataset directory when upgrading from an older downloaded
copy; a previously populated root can retain old metadata/video files. Pin the
repair commit for reproducible experiments. v3.0 names the dataset storage
format tag, not the Python package version. Use a LeRobot release supporting v3
(the validation environment is LeRobot 0.6.2 with the PyAV video backend).
Select the desired model inputs explicitly. A joint-space evaluation wrapper must be adapted to the absolute end-effector convention before simulation. Changing coordinate expression is different from physically changing a trajectory: the latter also requires IK/collision checks and consistent simulation rendering.
References
Source license: MIT. Upstream source files retain their notices.
- Downloads last month
- 428