Deploy Candidate Scoreboard

Master scoreboard

All checkpoints at model_2999. Cell color = good→bad for that metric's own direction. arm4 = low swing (~0.034 m foot), arm5 = high swing.

† arm5 s1 is a dead rear-leg-drag basin: its "survival" is stuck-but-upright, not a real gait. mechCoT measured only for the CoT-comparison subset. Deploy pick: arm5 seed2 — best falls (84), best curb (53.5), coolest (60.4), but thirstiest (mechCoT 1.21).

Fall-rate detail

Survival % across push magnitude (push_eval, Kp34/Kd0.8, 256 envs). The 3.0 column discriminates; ≤2.0 is near-100 everywhere.

seed2 checkpoint sweep

Why we don't just take model_2999: checkpoint quality is non-monotonic and invisible in TensorBoard (7/16 prior checkpoints didn't walk under the deterministic action). iter 2500 is a 30-point fall pothole; 2999 is genuinely best.

Gain sweeps seed2 @2999

Trained DR: Kp 22–34.5 / Kd 0.45–0.80. Deployed at Kp34/Kd0.8/grav-ff1.0 — the top edge, and sub-optimal on every count below.

Kp × Kd → fall survival @3.0 m/s

Lower Kd = more robust. Kp34/Kd0.7 is the optimum (89); the deployed Kd0.8 (82) is the worst damping.

Grav-ff × Kp → stance droop & torque (flat, Kd0.8)

★ = current deploy point. grav-ff cancels droop (more at low Kp) but the deployed Kp34/grav-ff1.0 over-shoots stance and spikes torque to 266 (vs ~185). Better: grav-ff 0.5.

Cost of transport

Mechanical CoT = ∫|τ·ω| / (m g d), the swing-sensitive energy term (stance holding is copper/τ², measured separately). Identical gains, flat, 1 m/s.

Key findings

• curbs are config-freeEvery seed of both arms climbs ~46–54%, despite arm4's 0.034 m swing. arm5's foot_clearance −2.5 bought zero curb capability — climbing is stance-push, not swing height.
• thermals are stance-dominatedAll seeds ~60–64 °C regardless of swing (calf heat = body weight through the knee, continuous). The lever is stance/knee-angle, not foot height. arm4 seed3's 86.7 °C is an asymmetric-load basin (loads one calf), not a swing effect.
• low swing wins efficiencyLow-swing arm4 is 27–42% cheaper in mechanical CoT — the efficiency win the low-swing idea was really about, via swing work, not thermals.
• the deploy tradearm5 seed2 = most robust + coolest but thirstiest; arm4 seed4 = ~27% longer runtime, same curbs/thermal, less shove-robust (70 vs 84%). Robustness vs runtime — a product call.
• MuJoCo terrace retractedIts thermal (63 vs 111 °C) and curb ("seed2 stuck") numbers were both foot-clipping artifacts. Only clean-Isaac (PhysX, frozen, mean-action) numbers survive.
• gainsKd 0.7 > 0.8 for falls; grav-ff 1.0 is over-gravved at Kp34. Deploy Kp34/Kd0.7/grav-ff0.5.

Multi-objective arm

The principled attempt to collapse the robustness-vs-efficiency frontier: arm4 base + efficiency (joint_power −6e-5) + thermal safety (torque_spread) + robustness (push ±2.0). Target: a seed that dominates seed2 on all axes — or an honest Pareto frontier if they trade.

mo1 (plain) is the recipe; seed1 stays the deployed pick. All five seeds walk, stay efficient (0.67–0.86), symmetric (L/R 1.01), and curb ~50% (a config property — every arm lands there). But the extra two seeds broke the "dominates every axis" story: falls now span 67–100% and Tss 58.9–64.5 °C. seed6 is excellent (0.67 CoT, 98% falls) but runs hot; seed5 is a genuine dud (67% falls, 64.5 °C). So the recipe reliably yields a good gait, but not every seed is deploy-grade — pick the seed (seed1/3/6 pass, seed5 fails). Consistent with this body's known seed/checkpoint variance. Symmetry, by contrast, is rock-solid across all five (L/R 1.01).

Mirror loss (mo1-sym) is a net negative at n=3. It's no more symmetric (both recipes ~1.00 in Isaac), less robust with a wider spread (mean 88 vs 96, seed3 down at 73%), and ~1.3 °C hotter (two seeds cross 60 °C at ~17 min); efficiency and curb are a wash. So it's dropped.

Why symmetry was never the lever. The asymmetry that motivated it (MuJoCo calf L/R 1.10) was a measurement artifact: the menagerie trunk carries the same +4.1 mm COM-y bias Isaac zeroes, and mujoco_walk_check wasn't removing it — so it measured a body the policy never trained on. On the symmetric body both recipes are L/R ≈ 1.00, so the mirror loss had nothing to fix. (Implementation aside: use_mirror_loss had been a silent no-op — the hand-copied CENet update() dropped the mirror-loss term; it was fixed and Loss/symmetry now converges, but the result stands.) The 3 MuJoCo harnesses now default to the symmetric body (--no-symmetrise for the raw sim2real asset).

Stair-mounting capability — the efficiency trade

The operator eyeballed that mo1 mounts tall steps less confidently than earlier arms. The d0.5 curb metric (0.14 m steps) can't see it — everything lands ~50%. A taller-step sweep on identical 0.23-range terrain does.

curb ≥3 %, climbing 3 steps of that height, identical 0.23-range terrain. Two arms were tried and both failed: the episode-40 s climb arm (no gain — curriculum time can't overcome it) and a joint_power-gated arm (mo1 with the efficiency penalty released on stepped terrain). The gated arm is in the table below.

The joint_power attribution was wrong. The gated arm was the single-variable test — releasing joint_power on stairs bought back no climbing (0.8% vs mo1's 4.7% at 0.23 m; both sims agree). So the −6e-5 penalty was not what suppressed mo1's climbing.

The soft-tracking-kernel hypothesis was also wrong. stand_spread runs a constant soft kernel (std 0.5) where mo1's Tall36Sharp anneals to 0.25 — the theory being that a sharp velocity-error penalty punishes the slow-down needed to mount a riser. The single-variable arm (mo1 + std 0.5) refuted it decisively: soft-kernel is ≤ mo1 at every difficulty (d0.5 43.4 vs 58.6, d0.8 23.4 vs 32.8, d1.0 3.5 ≈ 4.7) and nowhere near stand_spread. Softening the kernel just loosens gait discipline — it climbs no better and (operator-observed, borne out as a higher fall rate) twitches on missed footholds. MuJoCo agrees end-to-end, including at matched iteration 2500 (soft 0.8 vs mo1 1.1 steps). Flat tracking survived (95.6% at cmd 1.0), so the cost is climbing for nothing.

The push-magnitude hypothesis was wrong too — and it was the last one. mo1 raised the disturbance kick ±1.0 → ±2.0; stand_spread kept ±1.0. The single-variable arm (mo1 + push ±1.0, from scratch) did not recover the edge: d1.0 1.2% — below mo1 (3.5–4.7%) and nowhere near stand_spread (31–37%); d0.5 53.5 ≈ mo1, d0.8 21.1 < mo1. MuJoCo agrees (1.5 steps, inside the mo1 seed band 0.7–4.3). But it delivered the calmer gait it was also meant to: 3% MuJoCo falls, the lowest of the whole mo1 family (seed1 6%, seed5 20%), tracking held (94.8%). So ±1.0 is a real calmness/efficiency win — just not the climbing lever, and (see below) it is not a top-end robustness win.

Deploy axes vs the mo1 seeds — a Pareto trade, not a free win. Benched identically (Kp34/Kd0.8): push±1.0 is efficient and cool but collapses on hard shoves — falls@3.0 7% vs the mo1 band's 67–100. Training kicks at ±1.0 never taught it a 3 m/s shove (3× its envelope); its survival curve is 94/70/37/7 across 1.5/2.0/2.5/3.0. It wins efficiency (mechCoT 0.715, cheap end) and ties thermal (Tss 60.1 °C) and symmetry (L/R 1.00). So it sits in the calm/cheap corner: attractive only if the robot faces gentle perturbations, disqualifying if it faces 3 m/s shoves.

Verdict: stand_spread's climbing edge is not any single config knob. joint_power, kernel, and push were each ruled out alone (and the mo1 seed sweep shows climbing is heavily seed-dependent, 0.7–4.3 steps, none reaching stand_spread's 6.2). The remaining explanation is an interaction — or a seed-lucky basin. No more single-variable arms: the clean next step is a direct A/B of the two full configs at matched seed, or accepting the split — stand_spread the climber, mo1 the efficient/robust deploy pick — as a genuine Pareto frontier rather than one recipe that wins everything.

The pre-mo1 climbers pay for it: far less shove-proof (16–47% vs mo1's 96%) and hotter (fc05 is a 124 °C hot-calf disaster — the basin mo1's torque_spread was built to kill). So mo1's robustness+thermal are real wins; its climbing is the cost. All three single-variable arms (joint_power, kernel, push) are now spent without recovering the edge — the climbing gap is an interaction, not a knob. The remaining move is a full-config A/B or accepting the Pareto split (stand_spread climbs, mo1 deploys).

Generated from docs/RUNLOG.md and logs/ · clean-Isaac evals (push_eval falls, stair_eval curbs, isaac_thermal_table Tss, mujoco_walk_check mechCoT) · A1 DreamWaQ++ deploy campaign.