# v3 final dataset: VALIDATION

Validator: datasets/validate_v3.py. Full log: act/logs/validate_v3_run2.log.
Result: **ALL_CHECKS_PASSED** (26/26 checks OK).

## Dataset summary

- 500 episodes, 222451 frames, fps 50 TRUTHFUL (dt 0.02 s = every 2nd ctrl step of the 10 ms sim, k=3 slow replay).
- Duration: 7.87-13.27 s, mean 8.89 s.
- Schema: state (9,) qpos[0:9]; action (8,) absolute ctrl targets; 3 fisheye cams 480x480 (wrist_left, wrist_right, base).
- Teacher: measured-qpos slow replay, k=3, rim_clear=0.045, ik_speed=0.05, n_close=60, n_hold=40 (rl/slow/SLOWDOWN.md chosen config).

## Check bar (all OK, log lines in act/logs/validate_v3_run2.log)

1. Episode/frame counts: parquet rows 222451 == meta total_frames; 500 episodes contiguous 0..499; global index contiguous.
2. fps/dt truthful: timestamps start at 0, strict 0.02 s cadence in every episode (0 bad); meta fps == 50.
3. State/action ranges: state within joint limits + 0.05 rad soft tol (max excursion 0.0242 rad); action within actuator ctrl ranges + 0.05 rad soft tol (max excursion 0.0233 rad; MuJoCo clamps ctrl internally, ctrllimited=True on all 8 actuators). All finite.
4. Image integrity: decoded video frame counts == sum of episode lengths for all 3 cams; 10 random (episode, frame) spot-checks re-rendered from the full_qpos sidecar through the identical fisheye pipeline, MAE 1.39-3.05 (threshold 12), side-by-sides in datasets/v3/validation_pairs/.
5. Per-timestamp delta bounds: max |dq| <= 0.06 rad and max EE displacement <= 0.015 m between consecutive frames in ALL 500 episodes (worst observed: dq 0.0353 rad, dee 0.0116 m).
6. Contact == 0: kinematic re-simulation of every recorded frame of all 500 episodes from the qpos sidecars; ZERO robot-box contact frames anywhere; collection-time box_contact_steps == 0 for all kept episodes.
7. Task-label consistency: per episode, task string color == pinned target color == the one cube actually inside the box at the final frame (recomputed in sim). ZERO mismatches across 500 episodes. Exactly 5 task strings.
8. Gentle stats: all kept episodes passed the collection filter gentle AND correct AND box_contact_steps==0 AND duration in [5, 14] s (workers, act/logs/collect_v3slow_w[0-4].log); mean release vxy and peak EE speeds per color below.
9. LeRobot API: load + strided full-range iteration clean; meta matches reality.

## Per-color breakdown

| color | eps | frames | dur min/mean/max (s) | peak EE mean/max (m/s) | rel vxy mean/max (m/s) | contact steps | max dq (rad) | max dEE (m) |
|---|---|---|---|---|---|---|---|---|
| red | 100 | 44491 | 7.87/8.89/11.95 | 0.510/0.563 | 0.020/0.200 | 0 | 0.0343 | 0.0113 |
| green | 100 | 44038 | 7.93/8.80/12.28 | 0.510/0.579 | 0.018/0.146 | 0 | 0.0343 | 0.0116 |
| blue | 100 | 44367 | 7.96/8.87/11.71 | 0.516/0.580 | 0.033/0.499 | 0 | 0.0345 | 0.0116 |
| yellow | 100 | 44279 | 7.96/8.85/11.80 | 0.515/0.573 | 0.027/0.310 | 0 | 0.0353 | 0.0114 |
| purple | 100 | 45276 | 7.87/9.05/13.27 | 0.512/0.573 | 0.025/0.258 | 0 | 0.0353 | 0.0114 |

## Collection yield (per worker, one color each)

| color | kept | attempts | yield | rejections |
|---|---|---|---|---|
| red | 100 | 166 | 60.2% | {'not_gentle': 24, 'fast_not_gentle': 36, 'duration': 2, 'box_contact': 4} |
| green | 100 | 138 | 72.5% | {'not_gentle': 14, 'box_contact': 5, 'duration': 1, 'fast_not_gentle': 18} |
| blue | 100 | 147 | 68.0% | {'not_gentle': 26, 'fast_not_gentle': 20, 'box_contact': 1} |
| yellow | 100 | 154 | 64.9% | {'fast_not_gentle': 26, 'not_gentle': 21, 'box_contact': 7} |
| purple | 100 | 161 | 62.1% | {'fast_not_gentle': 32, 'not_gentle': 22, 'duration': 3, 'box_contact': 4} |

Total: 500 kept / 766 attempts (65.3% yield; SLOWDOWN.md predicted ~64%).