Correct bigstep card: invalid reward-term comparison -> measured m/s (0.009); attribute failure to foot-lift

u4_walk_bigstep_v1/README.md

@@ -1,4 +1,4 @@
-# u4_walk_bigstep_v1 — crouch "bigstep" walker (NEGATIVE RESULT)
+# u4_walk_bigstep_v1 — crouch "bigstep" walker (NEGATIVE RESULT: marches in place)
 
 Trained on skynet/SkyPilot, AWS us-east-2, 1x L40S (g6e.xlarge), storage-free.
 
@@ -12,34 +12,52 @@ Trained on skynet/SkyPilot, AWS us-east-2, 1x L40S (g6e.xlarge), storage-free.
 | Actions | 26 @ 50 Hz |
 | wandb | [`agioi2ya`](https://wandb.ai/sjohnu-eth-z-rich/rc_humanoid_rl/runs/agioi2ya) |
 
-## Result: bigstep did not improve on the crouch baseline
+## Result: it does not translate — 0.009 m/s
 
-| Walker | `track_lin_vel_xy` | `mean_reward` |
-|---|---|---|
-| **u4_walk_bigstep_v1** (this) | **0.778** | -1.77 |
-| `u4_walk_crouch_real_imudr` | 0.777 | -1.13 |
-| `crouch_nogait` | **1.838** | **+54.2** |
+Measured with `scripts/rsl_rl/measure_gait.py` (1000 steps x 64 envs, forward command forced).
+**Physical units.** Reward-term values are NOT comparable across configs with different
+weight/std -- see the correction note below.
 
-The bigstep envelope landed statistically indistinguishable from the crouch walker it was
-meant to beat (0.778 vs 0.777). `crouch_nogait` remains the best walker by ~2.4x on tracking
-velocity, and is the only one of the three with positive mean reward.
+| policy | cmd | fwd vel | cadence | step length | lean |
+|---|---|---|---|---|---|
+| r16_upright @ 0.8 | 0.8 | **0.634 m/s** | 3.30/s | 19.2 cm | 1.5° |
+| crouch_nogait | 0.5 | 0.572 | 8.14/s | 7.0 cm | 5.3° |
+| r16_upright @ 0.5 | 0.5 | 0.406 | 3.37/s | 12.0 cm | 0.7° |
+| u4_walk_crouch_real (22k) | 0.5 | 0.397 | 3.48/s | 11.4 cm | 9.9° |
+| r16_latency (10k) | 0.5 | 0.347 | 3.39/s | — | 13.3° |
+| **u4_walk_bigstep_v1 (this)** | 0.5 | **0.009** | 1.71/s | **0.5 cm** | 20.9° |
 
-## It converged by ~iter 7k
+It steps (1.71 steps/env/s) and reaches 0.122 m clearance, but translates 0.5 cm per step:
+it marches in place.
 
-| phase | tvel | ep_len |
-|---|---|---|
-| first 10% | 0.395 | 451 |
-| 25-35% | 0.745 | 694 |
-| 50-60% | 0.741 | 717 |
-| 75-85% | 0.780 | 787 |
-| last 10% | 0.778 | 711 |
+## Why: the foot-lift layer, not the latency
 
-Plateaued around 25-35% of training; the remaining ~18k iterations bought ~+4% tvel for
-~13 GPU-hours. **Cap future runs of this task at ~10k iters.**
+bigstep changed four things at once. The report-16 one-change-at-a-time series
+(`r16_*`, branch `test/SjohnU_remodel_actuators`) attributed the failure:
 
-Note: the negative `mean_reward` vs the early-training value is not policy degradation —
-early episodes are short (451 steps) and accumulate fewer penalties; reward drops as
-episodes lengthen to ~750.
+- **`r16_footlift`** (`feet_clearance` 1.0@0.15 -> 2.4@0.17, bigstep's foot-lift layer alone):
+  **0.001 m/s** -- the worst arm measured, worse than its own control. This is bigstep's killer.
+- **`r16_latency`** (~65 ms command latency alone): **0.347 m/s** -- latency HELPS. It was not
+  the culprit.
+- **`r16_latency_g200`** (latency + legs 200/10): 0.342 m/s -- the gain change is free.
+- **Lean is not causal**: `r16_latency` walks at 13.3° BACK, `r16_upright` at 0.7°, this fails
+  at 20.9°. Lean predicts nothing; the 20.9° here is a correlate, not the cause.
+
+The unifying rule (see `r16_upright`): the policy sits in a **stand-and-micro-correct local
+optimum**; anything that makes standing untenable (latency, spawn-tilt DR, more iters) forces a
+real gait, and anything that *enriches* standing (this foot-lift layer) kills it.
+
+## Correction (2026-07-16)
+
+An earlier version of this card compared `track_lin_vel_xy` reward-term values across policies
+(0.778 here vs 1.838 for `crouch_nogait`) and concluded `crouch_nogait` was "best by ~2.4x".
+**That comparison was invalid** -- those runs use different reward weights/std, so their
+reward-term values are not comparable. It also masked the real story: this policy scored
+tvel 0.778 while moving 0.009 m/s. Judge on `measure_gait.py` physical units only.
+
+The same card advised capping this lineage at ~10k because reward "plateaued by ~7k". That
+plateau was a *stuck* policy flatlining, not a converged walker -- `u4_walk_crouch_real` only
+learns to translate between 15k and 22k. Do not generalize this run's plateau.
 
 ## Files