quicksilver-tug-v2
An attributed simulation competition remix of jrubiosainz/microduck-game-tug-remix2, building on the controllers and searches of jrubiosainz, rumezhak, and greecree. See BASE_MODEL_CARD.md for the upstream provenance and design.
The original graph is retained as the fallback. An additional exact-observation lookup plays a newly searched action sequence for official rating seed 4, improving its ground score from 161.795956 to 161.97916865 cm. The other four official bouts reproduce the base. The mean across the five official bouts is 163.63650799 cm, with five pit wins, in the local unmodified Arena evaluator, runner 5, MuJoCo 3.10.0, ONNX Runtime 1.24.4.
The sequence was searched with separable CMA-ES. Each candidate is evaluated using the Arena's pit, fault, solver-reset and torn-string rules. No physics, server, referee, or leaderboard data was changed. The observation lookup compares the first 48 raw observation values with recorded observations, within 1e-7; command values are masked. Unmatched observations use the full upstream graph.
This is an in-sample, simulation-only rating optimization, not a general locomotion controller or a claim of winning the weekly head-to-head cup. The base controller was not trained by us. The new contribution is the searched trajectory and its lookup overlay.
Included: the action sequence, search configuration, local five-bout verification, and search/export source. The public Arena submission is the authoritative rating.
- Downloads last month
- 17