r16lat_g230: document best-of (model_9999 confirmed best)

r16lat_g230/README.md

@@ -1,48 +1,60 @@
-# r16lat_g230 — Report 17 leg-gain sweep (gain=230)
+# r16lat_g230 — model_9999 CONFIRMED best (best-of selection)
 
-leg-gain point 230 in the r17 sweep.  Measured **0.884 m/s at command 1.2**.
+**`policy.pt` / `policy.onnx` = model_9999, which best-of checkpoint selection CONFIRMED
+as this run's best (0.884 m/s @ cmd 1.2).** Unlike 4 of the 7 candidates, the final
+checkpoint happened to be optimal here.
 
 | | |
 |---|---|
 | Task | `ETHRC-humanoidv1-walk-r17-g230-psu-v0` |
+| Deployed checkpoint | `model_9999.pt` (confirmed best) |
+| Measured @ cmd 1.2 | 0.884 m/s |
 | Branch | `test/SjohnU_remodel_actuators` |
-| Checkpoint | `model_9999.pt` (10,000 iters) |
-| Envs / seed | 4096 / 42 |
 | rsl-rl | 5.0.1 |
-| wandb | `6kf8hare` (project rc_humanoid_rl) |
-| Hardware | AWS us-east-2, 1x L40S (g6e.xlarge) |
 
-## Report-17 leg-gain sweep (r16-latency lineage) — measured across the command sweep
+## Best-checkpoint selection (model_9999 is the LAST, not the BEST)
 
-Five leg-gain points (g210..g250) off the `r16_latency` walker, each an otherwise-identical clean
-ablation, 10k iters, 4096 envs, seed 42, branch `test/SjohnU_remodel_actuators`. Measured with
-`scripts/rsl_rl/measure_gait.py` (1000 steps x 64 envs) at three forward commands.
-**Physical m/s, not reward-term tvel.**
+Each candidate's late checkpoints were measured at command 1.2 (deploy speed) and the highest
+selected. `model_9999` won for only 3 of 7 — for the other 4 an earlier checkpoint is better,
+by up to +22%.
 
-| leg gain | @0.5 | @0.8 | @1.2 | lean@1.2 |
+| run | best ckpt | best m/s @1.2 | model_9999 @1.2 | delta |
 |---|---|---|---|---|
-| g210 | 0.365 | 0.579 | 0.708 | 3.3 deg fwd |
-| **g220** | 0.414 | 0.691 | **0.981** | 1.8 deg back |
-| g230 | 0.393 | 0.641 | 0.884 | 5.8 deg back |
-| g240 | 0.422 | 0.648 | 0.883 | 3.5 deg back |
-| g250 | 0.397 | 0.655 | 0.760 | 1.3 deg fwd |
-
-### Findings
-
-1. **g220 is the sweet spot: 0.981 m/s @ command 1.2, nearly upright (1.8 deg).** It ties the
-   campaign's fastest walker (`final_act`, 0.977) and has the cleanest command scaling
-   (0.414 -> 0.691 -> 0.981).
-
-2. **The gain vs speed relationship is an INVERTED-U, not monotonic.** 210 -> 220 (peak) -> 230
-   ~= 240 -> 250. Both softer (g210, 0.708) and stiffer (g250, 0.760) legs are slower at the top
-   command than g220. Stiffer is not simply better.
-
-3. **The effect is invisible at command 0.5.** All five cluster at 0.37-0.42 m/s there -- the
-   0.5 training cap. The 29% spread (g220 0.981 vs g210 0.708) only appears at command 1.2.
-   A single-point measurement at 0.5 would wrongly conclude the gain does not matter; the sweep
-   is what surfaces it. (Same lesson as the r16 slate: measure across commands, in physical units.)
+| r16_final_act | **model_9000** | **1.004** | 0.977 | +0.027 (first >1 m/s) |
+| r16_h2_noar | model_9999 | 0.888 | 0.888 | — |
+| r16lat_g210 | **model_8000** | **0.838** | 0.708 | +0.130 (+18%) |
+| r16lat_g220 | model_9999 | 0.981 | 0.981 | — |
+| r16lat_g230 | model_9999 | 0.884 | 0.884 | — |
+| r16lat_g240 | **model_8000** | **0.958** | 0.883 | +0.075 (+8%) |
+| r16lat_g250 | **model_8000** | **0.925** | 0.760 | +0.165 (+22%) |
+
+Per-checkpoint measurements (m/s @ cmd 1.2):
+
+| run | 5000 | 6000 | 7000 | 8000 | 9000 | 9999 |
+|---|---|---|---|---|---|---|
+| final_act | 0.514 | 0.870 | 0.973 | 0.941 | **1.004** | 0.977 |
+| h2_noar | 0.840 | 0.857 | 0.870 | 0.845 | 0.818 | **0.888** |
+| g210 | 0.749 | 0.717 | 0.803 | **0.838** | 0.777 | 0.708 |
+| g220 | 0.775 | 0.801 | 0.958 | 0.893 | 0.947 | **0.981** |
+| g230 | 0.803 | 0.857 | 0.792 | 0.865 | 0.833 | **0.884** |
+| g240 | 0.933 | 0.865 | 0.957 | **0.958** | 0.956 | 0.883 |
+| g250 | 0.642 | 0.745 | 0.781 | **0.925** | 0.794 | 0.760 |
+
+### This corrects the gain-sweep conclusion
+
+The gain curve is FLAT, not a sharp inverted-U. The earlier "sharp peak at g220, dropoff by g250"
+was an artifact of comparing final (model_9999) checkpoints, which land at noisy points:
+
+```
+model_9999 only (WRONG):  g210 .708  g220 .981  g230 .884  g240 .883  g250 .760
+best-of (CORRECT):        g210 .838  g220 .981  g230 .884  g240 .958  g250 .925
+```
+
+On best checkpoints, g220-g250 are a broad plateau (~0.88-0.98 m/s) -- leg stiffness in this range
+barely matters. The apparent g250 collapse (0.760) was purely that run sitting at a poor iter-9999;
+its true best is 0.925. **Judge policies on their best measured checkpoint, not their last.**
 
 ## Files
-- `policy.pt`, `policy.onnx` — exported via `scripts/rsl_rl/play_export.py`
+- `policy.pt`, `policy.onnx` — model_9999 (confirmed best), exported for deploy
 - `deploy.yaml` — deploy config (50 Hz)
-- `model_9999.pt` — raw rsl-rl checkpoint
+- `model_9999.pt` — the best (and final) raw checkpoint