djduecaster/dfreq-rollers

Roller declared-start specialist — local 2.570 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 2.570 s, 5/5 finishes, selected command 1.1. Individual declared-start times: [2.565, 2.680, 2.525, 2.535, 2.545] s. Validation used the full official sweep: 18 commands from 0.30 to 2.00 in 0.10 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. Seed 2 deliberately retains the earlier safe trajectory: transferring the later continuation to that start produced a forward fall at 1.76 seconds. Seeds 1, 3, 4 and 5 use the selected later continuation transfers. The rejected seed-2 result remains in the private research archive and is shown in the visual comparison.

Checks and limitations

All five selected teachers replay exactly. Five previously consumed starts 900001–900005 reproduce original N8 fallback exactly; these are not fresh held-out trials. All 770 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.

Some selected trajectories show increased lateral drift. Faster declared-start completion is not evidence of more robust general-purpose locomotion or physical-robot safety.

Provenance and licensing

Own fallback: djduecaster/dfreq-rollers@bc819769a2eea05078951f9befc95c481f480402, SHA256 dc660e8128c3ad2446a0cee9c2f8690ccae3dab681509c57010f91a59d78e797. The frozen N8 fallback derives from the Pollen Robotics Roller Sprint reference, revision c72fb92594e8b415487e87dfeed509141afc5e25, original model SHA256 01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a. Credit Pollen Robotics and the original N8 development lineage documented 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: 1,040,287 bytes; SHA256 79019a39d0e3d314db2b6a90c31f5575b70ccf6b235bc3acee5163aeaeedd61c. 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.