Muhammadsafi/microduck-sprint-cpu-remix

MicroDuck Sprint CPU Remix

A simulation-only remix of jrubiosainz/microduck-game-sprinter-remix2 for the Microduck Arena 2 m Sprint.

Official Arena result: 0.860 seconds, with all five official seeds finishing. Ranked #2 when verified on 17 October 2026 at 09:16 UTC. This ties the first-place source model; the earlier submission retains first place. Official run. The server score matches the local ARM64 result below.

Changes and search

This model starts from the source's pretrained network and ONNX step-counter wrapper. Additional joint/command offsets and sparse changes to the final network layer were searched locally. It was not trained from scratch.

The 30-minute search completed 123,480 candidate evaluations on an Apple M5 Max using six CPU workers. It tried CMA-ES on constant and time-dependent offsets, PGPE with ClipUp, and sparse elitist hill climbing on the final layer. The selected model came from the final-layer hill-climbing phase. The methods received unequal portions of the budget, so this is not a controlled comparison of algorithms.

Optimization used the five official seeds. The source's head-roll step-counter mechanism remains present. The Arena runner, scoring rules, and simulation were unchanged. All model inference used ONNX Runtime's CPUExecutionProvider; no paid cloud training was used for this remix.

Local evaluation

Runner 4 at Arena commit d7571179242140cefd788639a23195bd56575490, MuJoCo 3.10.0 and ONNX Runtime 1.24.4:

Official seed	Finish time
1	0.860 s
2	0.860 s
3	0.860 s
4	0.860 s
5	0.860 s
Mean	0.860 s

Two fresh-process evaluations, including the full 21-speed sweep, produced identical results. The selected command was 1.15 m/s. Raw results are in full-run-1.json and full-run-2.json.

After freezing the model, one predeclared check on seeds 1001 through 1050 found 20/50 finishes for this candidate and 17/50 for the original source. Both failed most of those seeds. This three-finish difference does not establish a reliable generalization improvement. These seeds were not used for optimization or candidate selection; raw outcomes and the plan are included.

Provenance and credits

Direct source: jrubiosainz/microduck-game-sprinter-remix2, revision f60a0155a464d4c4909cfd64f667afb4394a5a11, SHA256 5b2341d46805505b241e8031f7bc910af882d9351cd18f7966db4d2b0d8e1d8a.

Credit to jrubiosainz, suraj-phanindra, juicenv, namenu, and Pollen Robotics, following the ancestry documented by the source model. The source is a remix of suraj-phanindra/microduck-sprint-hc5-aifund, whose card credits earlier policies. This repository makes no claim to the original network training. The source declares no license; this repository does not add one.

Published policy.onnx SHA256: a08400cd3d3bc3720afb0aec1aeaf69e179248d3744aebf9ef9e0d7ed5e719ed. Input obs has shape [1, 61]; output actions has shape [1, 14].

Simulation only

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