dfreq-hill — reliable posture adapter, local 3.889 s

Simulation-only event-specialized controller. Publication is distinct from Arena submission; results below are local evaluation.

The unchanged current official runner5 evaluated all18 command-grid speeds on seed1, then the best three across five starts. Selected .60m/s: 3.920,3.865,3.915,3.870,3.875s (mean3.889), 5/5 finishes. The exact pinned own public baseline reproduced5.836s; improvement33.36%. This is not podium pace (observed third-place threshold .739s), and local results are not server results.

Own base: djduecaster/dfreq-hill@e070cbfbfb8a631a8065cc3c35b8eb1d73a634f7. Upstream Pollen reference: pollen-robotics/microduck-hill-climb-1m5-reference@d319208d92896353f581be754f6325d01f5e3732. All14 baseline initializer protobufs remain byte-identical. No neural training or fine-tuning. The inherited training field describes Pollen's training, not this search.

Change and scope

Eight fixed parameters transform forward command and projected gravity and add symmetric hip/knee/ankle, hiproll and neck/head residuals. Exact parameters are included in manifest.json. 281 parameter candidates, 562 bouts on two design seeds1,3. Memoryless61-observation→14-action ONNX; no runtime seed, clock, hidden simulation state or physics changes. Actual prior-action history retained.

The full command sweep includes failures at other commands; not an all-speed success claim. Zero or negative commanded velocity is still mapped positive by the inherited calibration, so idle/braking and real-robot safety are NOT validated.

Independent evidence

Exported ONNX matched Python adapter observations, actions and timed traces bit-for-bit across all five selected starts. The saved settle arrays were not included in that equality assertion. Twenty freshly chosen resets after optimization:20/20 finishes, mean3.89225s (3.825–4.075), versus baseline20/20 mean5.831s; faster in every paired case. local-FRESH20.json is the fresh-set evidence; local-HELDOUT.json overlaps historical tests and is only supplementary.

Visual diagnostic collision-geometry contact sheet inspected across all five official starts: upright walking, no tumbling; maximum tilt8.53degrees. Lateral displacement reaches.289m within.4m half-lane, so drift remains a limitation. Render omits visual body meshes and hides wall rendering only, never physics.

Reproduction and provenance

Official Space source revision f6072612b0e40ac613272f88818575b997e820b0; MuJoCo3.10.0, ONNX Runtime1.24.4, NumPy2.4.1. source-binding.json identifies a partial pinned evaluation-source restore; it is not a claim every website file was refreshed. Full research workspace retains source, meshes, trajectories, search rows and logs.

Pollen's upstream weights are credited. No standalone OSS model license is asserted. Prior attributed Hugging Face service-scoped remix history is not a blanket redistribution grant. This update preserves the existing attributed Hugging Face service-scoped derivative lineage and does not relicense inherited weights.

Expanded robustness check

100 additional untouched resets:100/100 finishes, mean 3.88560s, range 3.800–4.200s. Faster speed-only variants were not promoted: 2.868s official variant finished only89/100 on this same reset set.

This repository retains compact local evaluation evidence in files prefixed local-, exact deployed adapter parameters in manifest.json, upstream credit in upstream-base-card.md, and file hashes in SHA256SUMS. It is not a standalone research/runtime reproduction environment. No podium, physical-robot safety or all-speed reliability claim is made.