Rainbow Road TW3 BC

Recurrent behavior-cloning policy for Mario Kart Wii Rainbow Road.

This is the validated common-schema migration of the legacy clean100_joint_tw3/best policy. The original 172-observation encoder, GRU, control layers, and 17 demonstrated action rows are preserved. The checkpoint exposes the current 223-observation and 180-action contract; the 51 added observations enter through a zero-output residual encoder, and inference masks the action table to the 17 demonstrated item-free combinations.

Runtime contract

  • Funky Kong + Bowser Bike
  • 150cc, Manual drift
  • Rainbow Road, Solo Race
  • CPU Normal
  • Items None
  • Frame skip 4
  • Observation: rr-player-common-v2 (223 dimensions)
  • Action: v4-180-player-action-v1 with the 17-action BC mask

Validation

  • Migration equivalence: 512 recorded observations, 0 argmax mismatches
  • Maximum mapped-logit error: 8.5831e-06
  • Closed-loop smoke rollout: finished
  • Game frames: 13,373
  • Wall events: 1
  • Respawn events: 0

The separately fine-tuned three-epoch candidate was not adopted because it failed its matched closed-loop rollout. This repository contains the lossless epoch-0 migration.

Usage

With the matching kart-env source at commit c90118f:

uv run visual.py vlabki/rr-tw3-bc --watch
uv run eval.py vlabki/rr-tw3-bc

The checkpoint includes its model configuration, normalization statistics, reference, dataset manifest, runtime YAML, and migration audit.

Downloads last month
45
Safetensors
Model size
575k params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support