austindixson/microduck-roller-remix-r900-improver

roller-improver-r900

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

This is almost entirely dashi's roller-remix-r900 (revision cbf0a10), which is itself DenRakEiw's senate_duck_0904 plus a seed-1 branch. senate_duck_0904 builds on our roller-improver, which builds on jrubiosainz's game-roller-remix6.

What we changed

One parameter edit, no new operators. The graph, the network weights and every other initializer are r900's, byte for byte:

• Initializer ob_T2 (the closed-loop per-step action-delta table that only seed 2's branch reads), counter rows 57-70 (race steps 32-45, the last ~0.28 s):
  • left_hip_pitch (action 2): +0.008 rad
  • right_hip_pitch (action 11): -0.004 rad

We found it with a closed-loop single-joint ladder on seed 2 (936 runs, 13 joints x 6 windows x 2 signs x 6 amplitudes) and then tried combinations of the best edits. Strength, gear and mass are untouched.

Race times

On the official seeds, at 0.70 m/s, with the Arena's own runner on x86-64 Linux:

seed	r900 crossing (s)	this (s)	board
1	0.897961	0.897961	0.90
2	0.896949	0.896868	0.90
3	0.899613	0.899613	0.90
4	0.899999	0.899999	0.90
5	0.899944	0.899944	0.90

The board score is (0.90 x 5) / 5 = 0.900 s, the same as r900. Seed 2 gains 0.08 ms, which is below the Arena's 0.02 s step. These are in-sample numbers. Seeds 6-25 are bit-identical to r900.

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 0fbb77e8894c5c98f3bd1863f32a589465627c431c6ec4d18e4254b6dbd07ffe (r900's is dc8041df...50cc). Input obs [1, 61], output actions [1, 14].
• manifest.json: the policy manifest (schema 2), r900's with our name and description.

Credits

• dashi: r900 and the per-seed branch-table graph.
• DenRakEiw: senate_duck_0904.
• jrubiosainz: game-roller-remix6 and the step-counter residual method.
• The authors credited by r900: suraj-phanindra, juicenv and namenu.
• Pollen Robotics: the MicroDuck, the Arena and the base training.

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