Muhammadsafi/microduck-hill-cpu-remix

MicroDuck 1.5 m Hill Climb CPU Remix

A simulation-only remix of jrubiosainz/microduck-game-hillclimber-remix2 for the Microduck Arena 1.5 m Hill Climb.

Official Arena result: 0.780 seconds, all five official seeds finished. Ranked #2 when verified at 2026-10-17T09:55:39Z. This ties the source score; it is not a faster result. Official run. The server score matches both fresh local evaluations.

What changed

The pretrained network and the source's step counter are unchanged. This remix changes only two existing ONNX tables: joint-target offsets in OFF and the normalized forward-command column in C. There are 42 search parameters: three knots for each of 13 non-counter action channels and one normalized command channel. The interpolation and parameter values are included in recipe.json and parameters.npy.

CMA-ES evaluated 23,856 candidates in 1200.0 seconds using 2 CPU workers on an Apple M5 Max. The initial step sizes were [0.0005, 0.002, 0.006, 0.02, 0.05]. All inference used ONNX Runtime's CPUExecutionProvider. The Arena runner, event, scoring and five official seeds were unchanged. This was a local CPU search; no paid cloud training was used.

Candidate selection prioritized the official five-seed mean. A bounded extra term used remaining distance at a fixed frame from the source trajectory to distinguish candidates with equal finish times. That term is a search aid; it is not an Arena score or evidence of generalization. The selected candidate ties the source on the official scoring procedure.

Local verification

Two fresh processes ran the complete 18-command Arena evaluation and produced identical results at selected command 0.3 m/s. The runner was version 4 at commit d7571179242140cefd788639a23195bd56575490, with MuJoCo 3.10.0 and ONNX Runtime 1.24.4.

Official seed	Finish time
1	0.780 s
2	0.780 s
3	0.780 s
4	0.780 s
5	0.780 s
Mean	0.780 s

These are in-sample results: optimization used these same five seeds. No held-out evaluation was performed for this remix, and no improvement on unseen conditions is claimed. Raw fresh-process results are in local-evaluation-1.json and local-evaluation-2.json.

python rebuild.py reproduces policy.onnx byte for byte from the included source, parameters and recipe (numpy 2.4.1, onnx 1.22.0). A graph comparison confirmed that all content outside the C and OFF tables remains unchanged.

Provenance and credits

Direct source: jrubiosainz/microduck-game-hillclimber-remix2, revision 3f1b2f3d4b9c22b984eae26adcd3f8cee830e172. Source policy SHA256: 05b3d386bb5ba13f3e5cb53c027c0a5e5bd57308257728bd901a35405f98e104.

Credit to jrubiosainz, juicenv, suraj-phanindra, namenu and Pollen Robotics. The direct source credits juicenv/microduck-hill-es8, which descends from suraj-phanindra/microduck-hill-hc2-aifund and namenu/microduck-arena. The original training and wrapper design belong to their authors. This repository makes no claim to training the underlying network. The source declares no license; this repository does not add one.

Published policy SHA256: 8d33040a86fbccd0236e0cfce210874670398dc3656b18a771b3dcf0446c1d1c. Input obs: [1, 61]; output actions: [1, 14].

Simulation only

This controller has not been tested on a physical robot. Its action-encoded counter depends on the simulator feeding previous actions back exactly. Do not deploy it on a physical MicroDuck.