An audit of 4,940 robot episodes found a four-to-one idle-time gap inside one lab's data

The separation of measurement from judgement is the part I would copy. Give a diagnostic the power to drop episodes and you will spend the next sprint arguing about its threshold. In RDA v0.9.7 only L1's nine deterministic hard checks can set an EXCLUDE verdict; L2's eight motion and video metrics, L3's four profile metrics and L4's per-dataset P10/P50/P90 aggregation report and nothing more ^[4][5].

The two xArm sets are the cleanest comparison in the audit, because the confounds are held still: same platform, same lab, same team, 800 episodes each ^[6][7][8]. Median effective motion runs 79.2% on the lift task and 16.7% on the push task, a 4.7x spread ^[7][8][1]. Episodes carrying at least one L2 finding run 72 of 800 against 629 of 800, roughly nine times as many on equal samples ^[2]. The write-up attributes the gap to the tasks: lift is large-amplitude pick and place, push is a small-force nudge then watch ^[9]. Its warning about the consequence is stated as structure, not as a result: the post says a loss function trained on a 75%-idle distribution is structurally biased toward predicting inaction ^[10].

The discontinuity table is where I would be slowest to transfer a number. Spikes are found with MAD-based detection at default thresholds ^[11]. jaco_play flags 176 of 1,085 episodes at a median of 10 spikes per episode; libero_10 flags 0 of 379 at the same median of 10 ^[15][16]. Same median, different flag rate. MAD-based detection sets its bar from the dispersion of the sequence it is scanning, so I would treat the flag rate as a property of each dataset's action space as much as of its spike count. Both ALOHA scripted simulation sets flag every episode they have, at medians of 33 and 49 spikes, which the post puts down to scripted controller discretization ^[12].

Per-dataset episode counts appear for nine of the 13 sets. They sum to 3,555 episodes, or 72% of the 4,940 audited, leaving 1,385 episodes not broken out ^[3].

For the idle finding to bite on someone else's stack, the policy would have to see frames at close to the collection rate. Subsample, or predict action chunks, and the idle fraction reaching the loss is not the 83.3% measured on xarm_push_medium ^[8]. Every figure here is a measurement of a dataset under one tool's defaults, and the post does not report a training run; its own framing is narrower than a benchmark, saying the results challenge the assumption that "clean data" and "good training data" are the same thing ^[19].