microduck-hill-quicksilver-v2

A simulation competition remix of rumezhak/microduck-hill-remix-leader3, credited to rumezhak and the upstream authors identified in that model's card. The fallback network and all original graph operators are preserved. An additional observation lookup plays a newly searched joint-target sequence when its recorded observations match. Other observations use the original controller.

The new sequence was optimized locally with separable CMA-ES against the unmodified Arena evaluator, runner version 5, MuJoCo 3.10.0, ONNX Runtime 1.24.4, NumPy 2.4.1. Local five-start mean: 0.739 s, 5/5 finishes, command 0.15 m/s. The server's submission result is authoritative.

This matches the pinned base's 0.739 s official mean and improves our earlier quicksilver-hill-v1 entry (0.740 s). The new seed-1 trajectory gets closer to the 0.730 s timing step, but still finishes at 0.735 s; this is not an outright speed record over the credited base.

This is deliberately an in-sample, simulation-only competition policy. The lookup was optimized for the declared evaluation starts; these results are not a claim of general-purpose locomotion or safe real-robot operation. No evaluator, physics, leaderboard data, or server code was changed.

Reproduction artifacts: search_race.py, search-config.json, searched-actions.npz, and local-verification.json. The stored sequence can be replayed through the Arena's runner and overlaid on the pinned base with the supplied script's Recorder and overlay.

The search starting trajectory was first converted to equivalent servo targets, preserving every recorded motor force, the settle, and the full motion trace on the searched start. This removes unused target magnitude beyond motor saturation. The subsequent search changed the trajectory. The transformation code, verification, and starting action sequence are included.