Datasets:

---

ExylosAi

/

table_spill_cleanup_bimanual

like 2

Follow

Exylos 3

We are preparing larger EXYLOS releases beyond this 50-episode public preview.

The current preview is mainly for format inspection, loading tests, and small experiments. For the full release, we want to understand what would make the dataset actually useful for your robot learning pipeline.

What matters most for you?

• More episodes: 500 / 1,000 / 5,000 / 10,000+
• Format: LeRobot, Parquet + MP4, RLDS, HDF5, Zarr, ROS bag, WebDataset
• Modalities: RGB, multi-view video, wrist/egocentric view, depth, masks, object poses, object states
• Labels: success/failure, action phases, recovery behavior, collisions, object interactions
• Splits: train/validation/test, held-out evaluation scenes
• Task variation: object poses, distractors, lighting, camera angles, failure cases
• Baselines: loading script, visualization notebook, training config, evaluation script
• License: research preview, commercial training, custom task-family dataset

Please reply with your robot / model stack and the top 2-3 things you would need from a larger release.