The trained walker actually walks

[abgnydn]

The published Vaxenburg walking policy now walks the flybody under real physics with the kinematic assist off. Four port defects, each a silent disagreement with upstream flybody, each verified against upstream source.

Measured

tools/walkbench.mjs, kinematic assist explicitly off, 3 reps:

	before	after
capsized	every rep	never, 3/3
uprightness at end	−0.78 (on its back)	+0.997
speed	0.15 cm/sim s (tumbling)	2.004–2.021 cm/sim s
|action|max while upright	17.8–83.2	~5.9 (native band ~6)

Commanded speed was 2.0 cm/s — tracking within 1%. Control, policy disabled, assist off: 0.032 cm/sim s. The fly stands still, so the locomotion comes from the policy and nothing else.

The four defects

1. world_zaxis read the wrong axis of the rotation matrix — the third column where flybody reads the third row (fruitfly.py: MJCFFeature('xmat', root_body)[6:]). Transposes, so pitch and roll were sign-inverted, on what is by input-weight L2 norm the 2nd-highest-gain row of the input layer. The righting response added to a tilt. Independently confirmed against the shipped normalization stats: mean(accelerometer) ÷ mean(world_zaxis) = 981, exactly the gravity in fruitfly.xml.

2. The spawn height ignored the floor. floor.xml puts the plane at -0.15; _SPAWN_POS is calibrated against a plane at 0. The lowest claw started 0.15 cm up instead of 0.00048, so every episode opened with ~17.5 ms of free fall, claws off the ground and touch/force observations reading zero, exactly while the policy was establishing its gait.

3. 455 of 741 observation dims never read the body. Both branches differenced the reference against itself, so a tracking controller was told "zero error, perfectly aligned" on every tick, right through the capsize. Now Rᵀ_fly·(refPos − flyPos) and conj(q_fly) ⊗ refQuat against the live pose, indexed from f=0, over an absolute world trajectory the fly can fall behind.

4. flybody's actuator filter dynamics were missing. Upstream applies dyntype='filter' programmatically in fruitfly.py:_build, so it never reaches the shipped MJCF and a port reading the MJCF cannot see it. model.na was 0 — every control tick drove an actuator 100% to target instead of 18%, a plant ~5.5× stiffer than the one the policy was trained on, and the 59-dim actuator_activation observable was all zeros. Patched into the MJCF text in memory (the runtime reads it from the baked bundle, not from public/), asserted via model.na === 78 so a silent no-op throws.

Also

The kinematic assist is cleared under the policy again, restoring what a9e24f1 reverted. That revert was correct then — removing the assist from a broken walker made the feature visibly fail. It walks now, so the policy path no longer needs it. The CPG path still does, and every "assist off = pirouettes in place" measurement in LIMITATIONS §8 remains true for it and is now labelled as such.

The RL test gates on real locomotion instead of finiteness. Threshold 0.5 cm/sim s sits 4× below measured and 15× above the policy-disabled floor; the pre-fix value of −1.174 cm fails outright.

Not changed

• The connectome still does not walk the body. Brain→spine still yields a walking magnitude and turn bias scaling a hand-written tripod CPG; ~20.3M edges still reach the body as about one scalar per tick. What works is a published RL policy, not the fly's brain.
• The forward pass has still never been compared against the released SavedModel. verify_walking_policy.py imports no TensorFlow. The walker working is strong circumstantial evidence the architecture guess is right; it is not that comparison.

Verification

tsc clean · unit 5/5 · full e2e 34/34 on the baseline containing fixes 2 and 4. Two later full runs were contaminated by unrelated load on this machine (11 vitest workers from another session; 15-min load average 10.78; tests taking 6–8 min against a normal 40 s). Those runs failed 3 and 4 tests respectively with zero overlap, and all 7 pass individually on an idle machine — the signature of load, not regression. The RL test itself passed 6 separate single-test runs at 2.016–2.020 cm/sim s.

🤖 Generated with Claude Code