djduecaster/dfreq

Sprint declared-start specialist — local 0.848 s

Benchmark-specialized simulation policy. Local evaluation is distinct from Arena submission.

Verified local result

The pinned official Arena CPU evaluator, revision 76dc0e06a6d261bcf9bdf7cce3fdc062c3c14b23, runner v5, returned 0.848 s, 5/5 finishes, selected command 1.15. Individual declared-start times: [0.850, 0.850, 0.850, 0.850, 0.840] s. Validation used the full official sweep: 21 commands from 0.20 to 1.20 in 0.05 steps, then the official three-finalist/five-start procedure. Results are local, not an Arena server result or rank claim.

Method and specialization

This model combines our own optimized action trajectories with the original policy as fallback. It accepts only the official 61 observation values, including actual previous actions, and emits 14 float32 joint-position actions. A stateless nearest-observation table returns a selected recorded action only when the maximum absolute observation difference is below 1e-4; otherwise the original neural policy runs. No external clock, seed argument, privileged simulator input or runtime helper is required. Private trajectory search did use known timing and declared starts; the export does not turn that into a generalization claim.

All five starts were exposed to optimization/selection. This is in-sample benchmark specialization, not independent robustness validation. Physics, preparation, scoring and actuator limits were unchanged. Seeds 1, 2 and 4 use their respective remaining-start search winners; seed 3 uses the original 640-bout search winner; seed 5 retains its unchanged baseline trajectory. Later tie-break, torque-conditioning and radial diagnostics did not replace this artifact.

Checks and limitations

All five selected teachers replay exactly. Native fallback checks at unmatched commands 0.20, 0.80 and 1.20 reproduce the original policy (including its failures). All 339 saved observation rows retain exact selected outputs with tested input jitter up to 1e-5; tested offsets of 3e-4 fall back exactly. The exporter also tested 100 random unmatched observations. Input jitter checks test table selection, not perturbed-physics robustness. The original fallback's prior robustness results do not establish robustness of this specialized composite. Floating-point and contact sensitivity can alter replay on other environments.

The selected trajectories visibly tumble forward near the finish. They finish before the official 0.3-second fall-hold rejects them. This is valid benchmark completion, not stable sustained running, general walking, or physical-robot safety evidence.

Provenance and licensing

Own fallback: djduecaster/dfreq@7b67892a603eb95afb0ffbfe61d07b7d57128f12, SHA256 4aa8b27e4e02fbe9dcea1cae2d07f9505b51a62418dad853d2ff55e6a924a243. The inherited base retains suraj-phanindra/microduck-sprint-hc5-aifund. Its residual training lineage used sanitized outputs from jrubiosainz/microduck-game-sprinter-remix4. Credit: suraj-phanindra, juicenv, namenu, jrubiosainz, dashi, Muhammadsafi, and Pollen Robotics. The original base card is preserved in upstream-base-card.md.

No standalone upstream model license or blanket off-platform redistribution grant is asserted. Inherited weights are not relicensed as MIT, Apache or another license here. Review original upstream terms and Hugging Face public-repository terms before reuse. Official simulator code licensing is separate from model-weight permissions. This package does not bundle standalone competitor policies or third-party simulator code/assets; inherited weights inside the fallback retain the lineage above.

Artifact

policy.onnx: 1,239,639 bytes; SHA256 b0933952c5496945a9340b68bda25e256a3029a49540ab640b7730a275665109. Selected per-start source hashes are in selected-sources.json. Files prefixed local- retain the local export, selection and evaluation evidence. SHA256SUMS verifies these publication files. No independent fresh-validation or Arena rank claim is made. Simulation only.