dadudu/microduck-bubu-roller-so

bubu-roller-so

suraj-phanindra/microduck-roller-hc5-aifund (0.940 s) with its network unchanged, wrapped in a step counter that adds per-step offsets to its joint targets and to the forward command it reads, searched by CMA-ES on the Arena's official seeds: 0.932 s on those seeds. An entry for the 2 m Roller Sprint of the Microduck Arena. Simulation only — see below.

How it was changed

The source network is frozen: policy.onnx contains the source graph unchanged, with a few standard ONNX operators before and after it.

• step counter: the head_roll target is rounded to 1/256 rad and k/65536 rad is added, k the number of control steps run (saturating at 127). The Arena feeds the previous action back raw, so the next step reads k exactly. The first call (zero previous action) runs with k = 1; the 0.5 s settle is 25 calls, so the first race step is k = 26.
• offsets, zero before k = 26, piecewise linear in k on knots every 3 steps from 26 to 77, held after: one on the forward command the network reads (obs[48]), and one on each joint target but head_yaw and head_roll. 18 knots × 13 channels = 234 values.

The values were searched by CMA-ES (population 24, initial step 0.05, generation 1,467 of 1,500) on the event's five official seeds at 0.7 m/s, inside the Arena's runner (runner v4, MuJoCo 3.10.0, x86-64 CPU), with the fitness the board time plus a sub-step tie-breaker (the overshoot past the line over the run's mean speed).

Results

Race times (s) on the event's official seeds, at 0.7 m/s commanded forward speed (arena score, full sweep):

seed	microduck-roller-hc5-aifund	this remix
1	0.94	0.94
2	0.94	0.92
3	0.94	0.92
4	0.94	0.94
5	0.94	0.94
board (mean)	0.940	0.932

These are in-sample numbers: the search scored its candidates on these same seeds. The result is fragile: re-run 8 times with noise of σ = 10⁻⁷ on every observation but the counter's, it finished every seed every time, with boards of 0.936–0.940 s, never 0.932 s. A different CPU or library build could give the same.

Not for a real robot

Judged only in the Arena's simulator; nothing here was tried on hardware. The step counter rides on the head_roll target, which a real robot would not feed back bit for bit. Do not load it on a physical MicroDuck.

Credits

• suraj-phanindra for microduck-roller-hc5-aifund, the network used here, itself a fine-tune of juicenv/microduck-roller-es21, from namenu/microduck-arena.
• jrubiosainz's game-roller-remix2 for the step-counter start-offset idea, reimplemented here.
• Pollen Robotics: the MicroDuck, the Microduck Arena and its runner, and the publishing tools that wrote manifest.json.

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