siyuhsu's picture
describe the single-tar layout
5d9f1d7 verified
|
Raw
History Blame Contribute Delete
1.66 kB
metadata
license: mit

StellaVLA — RoboDojo trajectories

robodojo_lerobot.tar   120 GB   unpack in place -> robodojo_lerobot_v30/

Byte-identical copy of the joint variant of the RoboDojo LeRobot v3.0 export, kept next to the StellaVLA results so the trajectories and their annotations are guaranteed to line up. Contents of the tar:

path size notes
robodojo_lerobot_v30/meta/ 54 MB includes the modality.json upstream does not ship
robodojo_lerobot_v30/data/ 262 MB 14-D joint state + action parquet
robodojo_lerobot_v30/videos/observation.images.cam_high/ 48 GB head camera — what the CoT is grounded in
robodojo_lerobot_v30/videos/observation.images.cam_left_wrist/ 32 GB
robodojo_lerobot_v30/videos/observation.images.cam_right_wrist/ 33 GB

3500 episodes / 35 tasks / 1,859,602 frames, 25 fps, ARX X5 bimanual, 480x640. State and action are 14-D: [left_arm(6), left_gripper, right_arm(6), right_gripper].

Annotations (per-episode grounding chain-of-thought for exactly these episodes) are robodojo/robodojo_cot.tar in siyuhsu/StellaVLA-Data.

Upstream: RoboDojo-Benchmark/RoboDojo data/RoboDojo_lerobot_v30_video — unmodified. Benchmark: https://robodojo-benchmark.com (arXiv:2607.04434).

Do not substitute RoboDojo_ee_lerobot_v30_video. That variant is 16-D end-effector state and its episodes are split differently (ep1000 is 363 frames there vs 513 here), so its videos do not line up with the annotations frame for frame.

Code: https://github.com/siyuhsu/StellaVLA