Open LeRobot, RLDS, OpenX, HDF5, manifest-only, and adapter-backed datasets without rewriting your collection stack.
Open LeRobot, RLDS, OpenX, HDF5, manifest-only, and adapter-backed datasets without rewriting your collection stack.
The app is designed around existing robotics stacks. It should read the dataset shape your team already has, then produce reviewed references and QA output.
LeRobot-style folders and episode metadata.
RLDS/OpenX trajectory shards.
HDF5 bundles for ALOHA, ACT, RoboMimic, and generic profiles.
Manifest-only mode for sensitive datasets where raw media stays external.
Adapters normalize dataset metadata into the same episode, stream, action, state, and annotation model. The original data layout remains untouched.
{
"adapter": "lerobot",
"dataset_path": "./datasets/so101",
"episode_count": 128,
"streams": ["rgb", "depth", "state", "action"]
}Keep the clip, the cluster, the QA report, and the export manifest together so every training subset can explain itself.