RoboDojo demonstrations in the YAM LeRobot layout
The training demonstrations of the RoboDojo benchmark
(RoboDojo-Benchmark/RoboDojo at revision
269457a3b42d8d0ae69898a601f2c1487c2f826e), converted to LeRobot v2.1 datasets with three 168x224 camera streams
and 14-D state and action: the data layout of the TemporalRatio YAM policies. This is an unofficial conversion,
made for the TemporalRatio experiments, not by the RoboDojo authors.
| dataset | episodes | frames per second | source in RoboDojo-Benchmark/RoboDojo |
|---|---|---|---|
sim |
3,500 | 25 | data/RoboDojo_lerobot_v21_video (simulation) |
real_<task>, 18 tasks |
100 each | 25 or 30 (the robot's rate) | data/RoboDojo_real/<task>/<robot>/data/episode_*.hdf5 (real robots) |
19 datasets, 5,300 episodes, 3,467,943 frames, about 20 GB. Real-robot tasks: cap_pen, classify_objects,
cover_blocks, disassemble_LEGO, fill_pen_holder, hang_mugs, insert_charger, insert_tubes, make_bread,
make_food, pack_and_pour_fruit, pack_objects_into_backpack, put_objects_into_basket, stack_and_cover_blocks,
stack_bowls, stand_up_bottles, store_in_safe, sweep_blocks.
Features
| key | content |
|---|---|
observation.state |
float32[14]: left joints 1-6, left gripper, right joints 1-6, right gripper (joint positions in rad, gripper aperture as recorded) |
action |
float32[14], the same layout |
observation.images.left_image |
head (high) camera, AV1 video, 168x224 |
observation.images.left_wrist_left |
left wrist camera, AV1 video, 168x224 |
observation.images.right_wrist_left |
right wrist camera, AV1 video, 168x224 |
Each episode's task is its language instruction.
Changes from the source
- Frames resized from 480x640 to 168x224 with area interpolation and re-encoded as AV1 (libsvtav1, CRF 30, GOP 2, yuv420p).
- The real-robot JPEG frames hold BGR pixel data (a standard decoder swaps red and blue against RoboDojo's own previews); they are converted to RGB.
- One row per recorded step at the source frame rate. Actions are not resampled.
- State and action are the 14-D vectors above; for the real robots, the HDF5 groups
left_arm_joint_states,left_ee_joint_states,right_arm_joint_statesandright_ee_joint_states, concatenated. robodojo_conversion.jsonin each dataset records the conversion settings and, per episode, the source episode (its path in the source repo), robot and instruction.
The source repo has been updated since revision 269457a3; this conversion does not follow later changes.
Use
hf download imishutk0410/RoboDojo-YAM-LeRobot --repo-type dataset --local-dir $HF_LEROBOT_HOME/robodojo
The datasets then load with LeRobot (codebase v2.1) as robodojo/sim and robodojo/real_<task> under
$HF_LEROBOT_HOME.
License
Apache-2.0, the license of the source dataset (LICENSE). This is a modified version of the RoboDojo data; the changes are listed above.
Citation
Please cite RoboDojo:
@article{chen2026robodojo,
title={{RoboDojo}: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies},
author={Chen, Tianxing and Chen, Yue and Li, Zixuan and Tang, Junyuan and Su, Kailun and Wan, Weijie and Chen, Baijun and Lu, Haoran and Yan, Haowen and Su, Honghao and others},
journal={arXiv preprint arXiv:2607.04434},
year={2026}
}
- Downloads last month
- 104