Water current: completion and execution effort¶
Two tasks × three frozen DP checkpoints × three transverse currents × 30 fresh matched reset states: 540 full episodes. Controller gains, hydrodynamic coefficients, motor limits, rendering, action decoding and success criteria are fixed. Current is applied before camera warmup through robot-link relative-water forces. The zero-current baseline was newly executed for this study.
| Task | Current (m/s) | Seed 17 /30 | Seed 43 /30 | Seed 101 /30 | Mean ± SD (%) |
|---|---|---|---|---|---|
| OpenHatch | 0.00 | 30 | 30 | 30 | 100.0 ± 0.0 |
| OpenHatch | 0.10 | 30 | 30 | 30 | 100.0 ± 0.0 |
| OpenHatch | 0.20 | 30 | 30 | 30 | 100.0 ± 0.0 |
| RotateValve | 0.00 | 24 | 29 | 26 | 87.8 ± 8.4 |
| RotateValve | 0.10 | 19 | 20 | 19 | 64.4 ± 1.9 |
| RotateValve | 0.20 | 17 | 23 | 22 | 68.9 ± 10.7 |
Valve's paired differences from zero current are −23.3 pp (descriptive 95% interval [−43.3, −1.1]) and −18.9 pp ([−35.6, −2.2]). Intervals resample training and reset identities independently, keeping conditions paired, over 20,000 draws. They are marginal, not multiplicity-adjusted. The response is not monotonic.
Hatch retains full observed completion while 20-second motor-force RMS rises from 2.648 to 2.897 N at 0.20 m/s: +9.4%. This measures actuator effort, not electrical energy. All episodes supply the entire 20-second window; no allocation saturation is observed there. Valve contact-loss incidence is 1.1%, 36.7% and 46.7%. Contact loss requires prior contact and is an associated execution change, not a demonstrated cause of failure.
Figures, videos and evidence¶
The homepage Results → Policy × control → Water current contains interactive success, motor-effort and contact-loss graphs, with one line per training seed. Four silent, full-episode wrist recordings compare 0 and 0.20 m/s. Valve selects the first reset with baseline success/current failure; Hatch selects its first reset. Rerun outcomes match the scored outcomes in all four clips. The videos are illustrations and do not replace scored episodes.
- Frozen protocol
- Per-episode data
- Per-training data
- All paired effects and intervals
- Video provenance and selection
- Private evidence archive receipt
The intervention estimates response to the declared direction and onset, not a pre-equilibrated flow field. Mechanism dynamics are unchanged. A five-second expert pilot checks finite responses on disjoint development states; it is not a full-horizon expert success cohort under current.