jrubiosainz/microduck-game-roller-remix

MICRODUCK GAME roller remix

juicenv/microduck-roller-es21 (0.948 s), juicenv's entry, with Gaussian noise of σ = 3·10⁻⁵ added once to every weight and bias of its network: 0.944 s on the Arena's official seeds, 0.004 s faster than its source there. Same graph and command handling; simulation only.

An entry for the 2 m Roller Sprint of the Microduck Arena. Simulation only — see below.

What it is

A remix of juicenv/microduck-roller-es21 by juicenv: revision 2b4d5776, policy.onnx sha256 f56b4736285f…, 0.948 s on the Arena board (run 01M52N0XZ5QDMRQ3JN83Y5). Its manifest describes it as: 2 m Roller Sprint skater: namenu/microduck-arena (strata-roller-pb4, 0.960 s) fine-tuned by ES and an elitist hill climb on the Arena roller scorer, searched on x86 (0.948 s). Credit to namenu.; it names pollen-robotics/microduck_rl at 3c0598cae (branch hannes-ft) as the training code.

How it was changed

Only the network's numbers. Every weight and bias of its four Gemm layers (mlp.0.weight, mlp.0.bias, mlp.2.weight, mlp.2.bias, mlp.4.weight, mlp.4.bias, mlp.6.weight, mlp.6.bias; 197,774 values) got Gaussian noise of standard deviation 3·10⁻⁵: drawn in float64 from NumPy's default_rng([201, 1027]) (draw 1,027 of a lottery seeded 201), added, and rounded back to float32. Nothing else changed: the graph, its operators, its input normalizer, its ONNX metadata and its command handling are the source's. The file rebuilt from the source with that recipe is identical to this one.

The lottery drew such copies of the source at noise scales from 3·10⁻⁶ to 3·10⁻⁴, raced each on the event's five official seeds inside the Arena's runner (MuJoCo 3.10.0, x86-64 CPU), abandoned a copy as soon as it could no longer beat the source, and kept one only if it finished all five seeds with a lower board time. This draw was kept; it raced on its roller skates at the source's speed, 0.7 m/s forward.

Results

Race times (s) on the event's official seeds, at 0.7 m/s commanded forward speed, the speed both entries race at:

seed	microduck-roller-es21	this remix
1	0.96	0.96
2	0.94	0.94
3	0.94	0.94
4	0.96	0.94
5	0.94	0.94
board (mean)	0.948	0.944

These are in-sample numbers: the search scored its candidates on these same seeds, as the elitist searches that produced its source did. Times are whole control steps (0.02 s), so a board moves in steps of 0.004 s; nearby seeds or a different CPU can give other times.

Not for a real robot

Trained and judged only in the Arena's simulator, on its five official seeds; nothing here was tried on hardware. Do not load it on a physical MicroDuck.

Files

• policy.onnx: sha256 77438dee8c26a965d8992b2475f87e7c67344016810dbef08c40b9dacd0308ee, 793,832 bytes; input obs [1, 61], output actions [1, 14].
• manifest.json: Pollen's policy manifest (schema 2): a perpetual gait with a constant command, plus how it was made.
• README.md: this card.

Credits

• juicenv for juicenv/microduck-roller-es21, the policy remixed here, and everyone its card credits.
• namenu (namenu/microduck-arena): the source's ONNX metadata (run_path) points at namenu's strata training runs.
• Pollen Robotics: the MicroDuck, the Microduck Arena and its runner, the microduck_rl training code and the publishing tools that wrote manifest.json and this card's header.

The source repository declares no licence; this one declares none either and names the source as its base model.

Resumen (español)

Remezcla de juicenv/microduck-roller-es21 (juicenv) para el sprint de 2 m sobre patines de la Microduck Arena, solo para simulación: la misma red con un pequeño ruido gaussiano (σ = 3·10⁻⁵) sumado una vez a todos sus pesos, elegido por un sorteo elitista sobre el tiempo de carrera. 0.944 s en las cinco semillas oficiales, frente a 0.948 s de la original. Créditos: juicenv, namenu y Pollen Robotics.