r16_latency_g200/README.md

r16_latency_g200 — Report 16 one-change-at-a-time walker

Measured: the gains are free.:0.342 m/s — ✅ walks (= arm A)

A.2 — control + ~65 ms latency + leg gains 180/5→200/10. Paired with arm A, it isolates the gain change from the latency

Task	ETHRC-humanoidv1-walk-r16-latency-g200-psu-v0
Branch	test/SjohnU_remodel_actuators @ 8f9f058
Checkpoint	model_9999.pt (10,000 iters, completed)
Envs / seed	4096 / 42
rsl-rl	5.0.1 (native actor/critic state_dict)
Actions	26 @ 50 Hz
wandb	k28tg52k
Hardware	AWS us-east-2, 1x L40S (g6e.xlarge)

Report 16: one-change-at-a-time, measured in physical units

All arms = u4_walk_crouch_real + the r16 shared layer (25% standing envs, track_lin_vel_xy 1.5 -> 2.2) + exactly one further change. 10k iters, 4096 envs, seed 42, branch test/SjohnU_remodel_actuators @ 8f9f058.

Measured with scripts/rsl_rl/measure_gait.py (1000 steps x 64 envs, forward command forced). Physical units only. Reward-term values are NOT comparable across configs with different weight/std, and they actively mislead here: r16_footlift posts the best curves of the entire set (tvel 1.194, mean_reward 11.79) while moving 0.001 m/s.

policy	change	cmd	fwd m/s	cadence	lean	walks?
r16_upright @0.8	upright price + spawn-tilt DR	0.8	0.634	3.30/s	1.5°	✅ best
crouch_nogait	(ref)	0.5	0.572	8.14/s	5.3°	✅
r16_upright @0.5	upright price + spawn-tilt DR	0.5	0.406	3.37/s	0.7°	✅
u4_walk_crouch_real	(ref, 22k)	0.5	0.397	3.48/s	9.9°	✅
r16_latency	+~65 ms delay	0.5	0.347	3.39/s	13.3°	✅
r16_latency_g200	+~65 ms + legs 200/10	0.5	0.342	3.47/s	10.7°	✅
r16_control	— (baseline)	0.5	0.026	0.24/s	9.2°	❌
u4_walk_bigstep_v1	(ref, 4 changes, 25k)	0.5	0.009	1.71/s	20.9°	❌
r16_footlift	clearance 1.0@0.15 -> 2.4@0.17	0.5	0.001	0.28/s	11.9°	❌ worst

The rule: standing must be made untenable

The policy's default attractor is a stand-and-micro-correct local optimum — that is what r16_control is stuck in at 10k (0.026 m/s, 0.24 steps/env/s). Every result follows from whether a change makes standing tenable or not:

• Makes standing untenable -> a real gait emerges: latency (cannot twitch-correct at 50 Hz -> 0.347), spawn-tilt DR (must recover every reset -> 0.406), or simply training to 22k (crouch_real -> 0.397; it only learns to translate between 15k and 22k).
• Enriches standing -> it never walks: footlift (0.001, worse than its own control) and a slower gait clock (0.004).

Latency is load-bearing, not a deploy tax. control + latency walks at 0.347 m/s where the identical control marches at 0.026 -- same code, same 10k iters, one knob, 14x the stepping. Not a confound: the crouch-envelope chain already swaps legs_waist to IdealPDActuatorCfg and moves torque-speed into the PhysX spawn func; _apply_latency_only turns that into a DelayedPD with identical effort/velocity limits, gains and armature, and never touches spawn.func.

Leg gains 180/5 -> 200/10 are free (0.342 vs 0.347).

Foot-lift is what broke bigstep -- bigstep carried exactly this layer. It perversely reached lower clearance than control (0.1003 vs 0.1567 m) while paying 2.4x for height: it did not lift higher, it stopped stepping.

Lean is not causal. r16_latency walks at 13.3° BACK, r16_upright at 0.7°, bigstep fails at 20.9°. Lean predicts nothing -- though pricing it still helps, via the spawn-tilt/untenable -standing mechanism rather than via the lean itself.

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 (re-export / resume)