vlabki/rr-v2

RR v2

Rainbow Road 150cc bike policy: preserved v2 update 2500, with unchanged driving weights. Four individually competing local racers against Normal CPU: Funky Kong / Bowser Bike, Daisy / Mach Bike, Baby Daisy / Bullet Bike, Baby Luigi / Bullet Bike.

Items are ON by default (Recommended). The full heuristic item controller is enabled, including rear defense, POW response and conditional post-impact SSMT. team_mode: false: other local racers are opponents. Item use is not learned by PPO. Start and Lakitu recovery boosts are enabled.

Run

In a configured kart-env / Dolphin installation using the matching runtime source:

python visual.py vlabki/rr-v2 --env_id 0 -o rr-v2.png --temp 0

The bare model name is rr-v2 with the default vlabki namespace. No extra item flag or external config is needed. Checkpoint config.yaml includes the complete visual configuration. Explicit CLI game options must match the four-player Solo Race setup. --temp 0 matches the evaluation; the Discord bot otherwise defaults to temperature 0.3 and five races.

Runtime compatibility

Use the accompanying runtime-src.tar.gz in a separate configured kart-env checkout (or update the bot worktree to the equivalent source). It contains visual.py, Python sources, dependency files and a default runtime config. The matching source includes profile adapters and item fixes; weights alone do not update an older Discord bot's code. release.json records source and weight hashes. The archive excludes game images, Dolphin binaries, credentials and training logs.

Evaluation

96 item-on races, four model-controlled racers per race, temperature 0:

• Finish rate: 376/384 (97.92%).
• Mean finished three-lap race: 3:04.81; median: 3:02.93.
• The local racers occupied first and second in all 96 races; this is a roster-level result, not a 100% individual win rate or cooperative policy.
• Per-profile results are in evaluation.json. These figures use the same heuristic controller packaged here, and do not measure the controller's benefit in isolation.

Runtime weights, normalization, route reference and checksummed portable fallbacks are included. Optimizer/resume state is excluded.

Visual smoke test

One local bot-mode race at the default temperature 0.3 finished 4/4 in positions 2, 1, 4, 3; item use, POW inputs and post-impact SSMT inputs were recorded. PNG generation and 52 targeted tests passed. See visual-validation.json and visual-check.png. This is a local CLI test, not a live Discord bot deployment.

Matching Git source: train_RR commit 41a5fd0.