Record the configured run

Set the episode count here. Start recording is on Fault Injection, after the faults are configured. Each episode stores joint state, the commanded action, and a fault tag that is 1 only while the cube is falling. Each run is uploaded to its own folder. Each episode includes a wrist video and an agent video.

Drop is the fault this plugin runs. The other faults are not available yet.

Paste the S3 URI of your recorded dataset, then click Validate dataset before training.

Test a checkpoint on held-out episodes

75% of the episodes are scheduled to drop, once, at a uniform fraction of the pickup-to-basket distance. The rest are clean. A drop counts only if the gripper actually opens. The scene is the LIBERO tabletop.

Run validation to fill these scores.

Precision is the fraction of episodes the model flagged that really had a drop. Recall is the fraction of episodes with a drop that the model flagged. F1 is the harmonic mean of precision and recall. It is low when either one is low. A zero denominator is undefined, not zero. The label is whether the gripper opened. The prediction is whether any frame score reached the threshold.

This run has no clean episode, so a false positive cannot occur.

Recovery rate is the fraction of episodes where a drop fired and the object finished inside the basket. It ignores whether the fault was detected. Episodes with no drop are left out. This is not a precision, recall, or F1.

Pick rate: fraction of episodes where the cube was lifted >5 cm above its rest position. Place rate: fraction of episodes where the cube ended inside the basket. Place rate ≤ pick rate because you must pick before you can place.

—