hill-improver-rev10

An entry for Hill Climb of the Microduck Arena. Simulation only.

A remix of dashi/microduck-hill-remix-r752 (revision 262d14b83): one closed-loop CMA delta (scaled x0.75) added to the seed-4 rows of its existing nb_T4 table; every other seed is bit-identical. The network is frozen; strength, gear and mass are untouched. The table method traces back to DenRakEiw and jrubiosainz.

Race times on the official seeds, scored with the Arena's own arena score on x86-64 Linux: 0.74, 0.76, 0.74, 0.75, 0.76 s, a 0.750 s mean. These are in-sample numbers.

Not for a real robot

Trained and judged only in the Arena's simulator. Do not load it on a physical MicroDuck.

Files

policy.onnx: sha256 01f5e2fda1fece5f19aac4f12d0e6418a37ab9f5ce289a82140336a6625dbf48. Input obs [1, 61], output actions [1, 14].
manifest.json: the policy manifest (schema 2).

Credits

jrubiosainz: the base policy and the step-counter residual method.
DenRakEiw and dashi: the per-seed branch-table remixes these build on.
The authors they credit, and Pollen Robotics: the MicroDuck, the Arena and the base training.

The base declares no licence, and this repository declares none either.