angledusgar/forge-failure-v1

Forge failure v1 — the episodes the expert lost

The failed episodes from the same collection that produced angledusgar/forge-v1: identical policies, identical environments, identical recording — only the outcome differs. Same schema, so the two sets can be mixed or contrasted directly.

They are published apart from the successes on purpose. Imitation learning on failed trajectories teaches failing, so a behaviour-cloning consumer should never receive them by accident; a failure detector, on the other hand, cannot be trained or calibrated without them.

Contents

episodes	51
frames	13,899
tasks	all three Forge tasks, by per-frame task string
format	LeRobot v2.1, video (AV1), fps 15

Failures are rare because the experts are good: roughly 4% for PegInsert and 1% for GearMesh and NutThread, out of ~1,000 episodes each. Collecting 100 failures per task at those rates would need several thousand episodes per task.

These are tail events, not a representative failure distribution. They are what a near-ceiling policy happens to get wrong, which is not the same as the failure modes an intervention system is usually designed against. For a set where failure is common and mechanically meaningful — jams from a peg cocked in the gripper — the task has to be made harder at collection time rather than sampled from a policy that rarely fails.

Features

Identical to forge-v1:

key	dtype	shape
image, wrist_image	video (AV1)	256×256×3
state	float32	(8,) — eef pos (3) + axis-angle (3) + gripper (2)
actions	float32	(7,)
ft	float32	(6,) — force (3) + torque (3), EE frame, smoothed
ft_noisy	float32	(3,) — the force the policy observed

The wrench channels are the point of this set: on a 0.114 mm clearance the difference between approaching, contacting and binding is not visible at 224×224, but it is plain in the force trace.

Provenance and caveats

Collection recipe, expert checkpoints, and the PegInsert stock-override caveat are documented in the forge-v1 card and apply unchanged here.

Two notes specific to this set:

• Per-episode success labels are implicit. Every episode here failed; there is no success column, since a mixed set would need one and this one does not.
• The first collection run lost its in-memory failures. Failures were buffered and only flushed once enough accumulated, and that run was killed under memory pressure before the flush. The count here reflects what survived; roughly 19 failures from the first partial run were lost.