Galaxea Open-World Dataset taken from OpenGalaxea/Galaxea-Open-World-Dataset,
converted to LeRobot Datasets v3.0 format using lerobot.datasets.v30.convert_dataset_v21_to_v30.
Missing subsets
The subset Boil_The_Water_20250714_006 is missing due to the original files having some episodes at 62 fps,
which causes the conversion script to crash with an error.
The subset Put_The_Items_Into_The_Storage_Box_20250929_002_007 is missing due to it having 7 DoF arms rather than 6 DoF.
Approximate no-torso-movement subsets
The /no-torso-subsets.txt file contains a list of 110 subsets that mostly exclude torso movement.
There are still some torso movements included nonetheless. The subsets are selected by the criterion
\(\lVert \text{subset-level torso velocity mean} \rVert _1 + \lVert \text{subset-level torso velocity standard deviation} \rVert _1 < 0.03\).
The 0.03 threshold was chosen by trial-and-error with manual inspection / eyeballing of subsets near those values.
Dirty episodes
Among the no-torso-movement subsets, some subsets contain episodes where we encountered errors loading the corresponding video.
To document this so that those episodes may be skipped during training,
we added a meta/removed_episodes.json file to each affected subset's directory.
We did not check outside of the no-torso-movement subsets.
Task config
The no-torso-movement subsets also each have a task_config.json file with two fields:
"coarse_task_as_main_task": whether the subset's coarse task is deemed good enough to replace the fine task as the main task, and have the fine task be subtasks instead."rename_coarse_task": when used as the main task, the coarse task text should be replaced with this text if notnull. Ifnull, make no change.
Arm-and-gripper-only subsets
The /arm-and-gripper-only-subsets.txt file contains a list of 76 subsets that additionally mostly exclude chassis movement.
There are still some chassis movements included nonetheless. The subsets are selected by the criterion:
\(\lVert \text{subset-level chassis velocity mean} \rVert _1 + \lVert \text{subset-level chassis velocity standard deviation} \rVert _1 < 0.03\).
This 0.03 threshold was also chosen by trial-and-error with manual inspection / eyeballing of subsets near those values.