r16lat_g260lock — BEST-checkpoint bundle (model_6000), locked-arm gain 16
260
policy.pt/policy.onnx = model_6000, the BEST checkpoint (0.908 m/s @ cmd 1.2),
NOT the final model_9999 (0.736 m/s). model_9999.pt retained for reference.
| Task | `ETHRC-humanoidv1-walk-r17-g16 |
| 260-lock-psu-v0` (locked arms) | |
| Deployed checkpoint | model_6000.pt (best of the run) |
| Measured @ cmd 1.2 | 0.908 m/s |
| Branch | test/SjohnU_remodel_actuators |
| rsl-rl | 5.0.1 |
Report-17 WAVE 2 — locked-arm leg-gain sweep (g260-g300), best-of checkpoints
Five locked-arm gain points off the r16-latency walker, 10k iters, 4096 envs, seed 42, branch
test/SjohnU_remodel_actuators. Each policy is its BEST checkpoint by measured velocity at
command 1.2 (best-of selection), NOT model_9999 -- which was best for NONE of these 5.
| gain (locked-arm) | best ckpt | m/s @1.2 | model_9999 @1.2 |
|---|---|---|---|
| g260 | model_6000 | 0.908 | 0.736 |
| g270 | model_6000 | 0.874 | 0.686 |
| g280 | model_9000 | 0.937 | 0.664 |
| g290 | model_7000 | 0.866 | 0.705 |
| g300 | model_8000 | 0.879 | 0.777 |
Within the locked-arm sweep, 260-300 is a plateau (~0.87-0.94 m/s), g280 marginally best. No strong gain trend in this range.
CAVEAT: do NOT read g210->g300 as one continuous curve
Wave 1 (g210-g250) trained with FREE arms (...-g{210..250}-psu-v0); Wave 2 (g260-g300) trained
with LOCKED arms (...-g{260..300}-lock-psu-v0). The variants differ AND the gain ranges do not
overlap, so a cross-wave comparison confounds leg gain with arm-locking. Compare within a wave only:
FREE-arm (wave 1, best-of): g210 .838 g220 .981 g230 .884 g240 .958 g250 .925
LOCKED-arm(wave 2, best-of): g260 .908 g270 .874 g280 .937 g290 .866 g300 .879
Free-arm tops out higher (g220 0.981) than locked-arm (g280 0.937), but that difference is the arm-locking + higher gain range combined, not gain alone. A clean gain-vs-lock ablation would need matched gains with lock as the only toggle.
model_9999 is almost never the best
Across all 10 gain arms (both waves), the final checkpoint was best only ~3 times. In wave 2 it was best 0 of 5. Deploy the best measured checkpoint, not the last.
Files
policy.pt,policy.onnx— the BEST checkpoint (model_6000)deploy.yaml— deploy config (50 Hz)model_6000.pt— best raw checkpoint ·model_9999.pt— final, reference