r16_h_ankle — Report 16 SKY SLATE

H-ANKLE — the REAL measured ankle (rs_kp 8.0/1.5). WALKS: 0.540 m/s @ cmd 1.2. Resolves the deploy-critical ankle unknown positively.


Branch	`test/SjohnU_remodel_actuators`
Checkpoint	`model_9999.pt` (10,000 iters, completed)
Envs / seed	4096 / 42
rsl-rl	5.0.1 (native actor/critic state_dict)
wandb	`ijog4us3` (project rc_humanoid_rl)
Hardware	AWS us-east-2, 1x L40S (g6e.xlarge)

Report-16 SKY SLATE — measured across the command sweep (physical units)

All arms branch off H (r16_deploy) + exactly one change, 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 — tvel lied repeatedly (e.g. footlift scored tvel 1.19 at 0.001 m/s).

arm	change vs H	@0.5	@0.8	@1.2	lean@1.2	walks?
`r16_final_act`	action smoothing (action_rate -0.3)	0.407	0.697	0.977	0.2°	✅ fastest
`r16_h2_noar`	action_rate -0.3 -> -0.12	0.388	0.644	0.888	1.6°	✅
`r16_h_ankle`	ankle rs_kp 2.0/0.15 -> 8.0/1.5 (REAL ankle)	0.222	0.382	0.540	3.4°	✅
`r16_h_cmd`	limit_ranges lin_vel_y (-1.0,0.5)->(-0.2,1.2)	0.090	-0.035	-0.630	—	❌
`r16_final_phys`	action_rate -0.12 + 100x soft_landing/10x joint_acc	—	—	—	—	🔴 suicidal

What the slate establishes

The command cap WAS the speed ceiling. Every prior walker measured ~0.4 m/s because limit_ranges capped the forward command at 0.5. Given a higher command, the good walkers scale cleanly: final_act 0.41 -> 0.70 -> 0.98 m/s. final_act @1.2 (0.977 m/s, 0.2 deg lean) is the FASTEST walker of the whole campaign, beating upright (0.634) and crouch_nogait (0.572).
But widening the command range DURING training (h_cmd) backfired. h_cmd opened the range to (-0.2, 1.2) and never learned to walk forward — 0.090 m/s at 0.5, and it goes BACKWARD (-0.63 m/s) when commanded +1.2. Read together with (1): the right move is to train with the 0.5 cap and command higher AT DEPLOY, not to widen the training range.
The real measured ankle walks (h_ankle). Every prior walker used a PLACEHOLDER ankle; the only policy with the real ankle was bigstep, which couldn't walk -- a confound that sat between sim and hardware. h_ankle resolves it: the real rs_kp 8.0/1.5 ankle walks (0.222/0.382/0.540 m/s), at ~55% of final_act's speed. The deploy story holds.
action_rate -0.3 vs -0.12 is not decisive; the PHYSICAL-penalty overshoot is fatal. final_act (-0.3) and h2_noar (-0.12) both walk well, with -0.3 marginally faster. The exploration-tax audit predicted -0.12 would clearly win; measured, it did not. What IS fatal is final_phys: keeping action_rate at -0.12 but raising soft_landing 100x (-1e-4 -> -0.01) and joint_acc 10x made LIVING net-negative, so the policy optimized by terminating early (reward rose -24.7 -> -18.2 as episode length fell 46 -> 22). A textbook suicidal policy.

Files

policy.pt, policy.onnx — exported via scripts/rsl_rl/play_export.py
deploy.yaml — deploy config (obs/action layout, gains, 50 Hz)
model_9999.pt — raw rsl-rl checkpoint