flail-robotics/flail-episodes

FLAIL episodes

Every attempt a FLAIL robot makes, successful or not, in LeRobot v3.0 format.

Robots are spawned by anyone, run in a physics simulator, and are given tasks they mostly fail. The failures are the point: a world action model needs to have seen an action not work in order to learn what working looks like. Nothing here is curated for success, and nothing is thrown away for being bad.

Why this exists

Tesla, Figure and Physical Intelligence are collecting robot experience behind closed doors. Nobody outside those companies will ever see what went into those models. This dataset is the same kind of experience, collected in public, released as it accumulates.

Format

LeRobot v3.0, the same shape as OpenMOSS-Team/libero-lerobot-v3.0, so it drops straight into EasyWAM and anything else that reads the format.

Provenance

Each episode carries the hash that was written onchain when it was accepted. The accepting transaction is public, so for any episode in here you can check that it was submitted when it claims to have been, by the robot it claims to belong to, and that nothing has been edited since. That check does not depend on trusting this repository.

Rungs

Each rung trains on a frozen snapshot of this dataset. When a rung ships, its weights are released with the exact snapshot hash it was trained on, so a model card can prove which episodes went in.

rung	backbone	status
twitch	Cosmos 2B LoRA	not trained
reflex	Cosmos 2B full	not trained
grasp	FLUX.2 4B	not trained
reach	Wan2.2 5B	not trained
hands	Wan2.2 5B multi-task	not trained

Episode-hours per rung are not published, because they are not known. The ratio comes out of the first training run by measuring eval score against dataset size. Any number stated before that run would be invented.

Status

Collection has not started. This repository exists so the format and the provenance scheme are public before the first episode lands, not after.