djduecaster/dfreq-sock-heist

Sock Heist — green-phase neck/head posture update

Local official score 22.081 s, 5/5 finishes, external command 1.15 m/s, versus prior published governor 22.183 s: a 0.102-second speed improvement. Individual local finish times: 21.840, 23.370, 21.750, 21.485, 21.960 s. These are local results, not a new Arena server receipt or rank claim.

Actual change

The licensed Pollen walker and embedded normalizer remain frozen. Our existing turn-speed governor maps internal forward command by the delivered yaw-rate command. This version adds −0.075 rad to neck and head-pitch actions (indices 5 and 6) only when delivered forward-command observation 48 is nonzero: the event's green phase. It adds no residual during red/zero-command phases. The gate uses the legal delivered observation, not a private timetable or clock. The exact float32 graph and parameters are in export_selected.py and adapter-binding.json.

The model accepts the unchanged 61 raw float32 observations and emits 14 joint-position offsets at 50 Hz. Actual previous-action feedback stays truthful. It contains no seed table, observation-fingerprint lookup, hidden scoring clock or privileged simulator inputs. Official physics, delays, preparation and scoring are unchanged. No new neural training occurred; this is bounded adapter search.

Validation and limitations

The full official command sweep and exported ONNX evaluation used source revision f6072612b0e40ac613272f88818575b997e820b0, runner v5, MuJoCo 3.10.0, ONNX Runtime 1.24.4 and NumPy 2.4.1. Selected exported actions were asserted bitwise equal to the Python adapter on every action of five native replays; 200 random green/red observation checks had zero forward error. Side/overhead diagnostic collision-geometry sheets were inspected privately: walking/stopping and posture/yaw swings, no collapse on these five official finishes. Visual inspection does not establish contact-force causality or physical-robot safety.

No robustness improvement is claimed. Across the combined 100 broader resets, both old and new models finished 89/100. The first 20 starts were already used for screening. On the later fresh subset (starts 21–100), the new model finished 69/80, versus 70/80 for the old model. Combined 40-second-failure-penalized means were 24.07510 s new and 24.39785 s old. The new model was selected after comparing these candidates, so these results must not be represented as an untouched final validation set for a later selection. Falls and caught-by-the-giant failures remain. A faster yaw-only contender was not preferred because combined reliability was worse.

The original uncorrected Python screen had float32 association/parity errors and is excluded from accepted evidence. Trusted screens followed the parity repair. Official starts were exposed during selection; this is an event-specific simulation policy, not a generally safe walker. No podium claim or real-robot braking/stability guarantee is made.

Provenance and license

Inherited alpha_walking.onnx is from Pollen Robotics microduck-policies, revision 665ae4eb20bfe13bca1ef4ea6d4bb6b4a096e448, 793705 bytes, SHA256 e36332d383997d51401897734cd3e79cf5038406feddb18b4d57ecfb141daa6c. This exact model's upstream card declares Apache-2.0. LICENSE and upstream-license-verification.json retain that verification. Credit remains with Pollen Robotics and original model contributors; no endorsement is implied. NOTICE and policy.onnx.NOTICE.txt identify our modifications.

Previous own governor: djduecaster/dfreq-sock-heist@f1346106465efafb8b8ada1bedad5395d7794769, SHA256 522f37107cb5467a1bb414e5ec5e04bada1c555856a4d11751cafef5df018298. Existing baseline/old export receipts remain historical evidence; current evaluation is evaluation.json and the files prefixed local-. No meshes or separate competitor policies are included.

Current policy.onnx SHA256: e9300f63f8374029ea21c732f78f27a84b54926baabe0d240d13cd27a669f44a. SHA256SUMS covers the updated publication files. Simulation only.