---
license: mit
library_name: stable-baselines3
tags:
  - BVS
  - bvs-schema-4
  - reinforcement-learning
  - onnx
model-index:
  - name: BVS-default
    results:
      - task:
          type: reinforcement-learning
        dataset:
          name: cpu easy
          type: bvs-matches
        metrics:
          - type: win_rate
            value: 1.0000
            name: win rate versus cpu easy over 20 matches
      - task:
          type: reinforcement-learning
        dataset:
          name: cpu hard
          type: bvs-matches
        metrics:
          - type: win_rate
            value: 1.0000
            name: win rate versus cpu hard over 20 matches
      - task:
          type: reinforcement-learning
        dataset:
          name: cpu normal
          type: bvs-matches
        metrics:
          - type: win_rate
            value: 1.0000
            name: win rate versus cpu normal over 20 matches
---

# BVS default

A fighter policy for Bro vs Sis, trained with PPO (stable-baselines3) and exported to ONNX.

## Contract (`bvs-schema-4`)

- MLP only: the graph may use `Gemm`, `MatMul`, `Add`, `Mul`, `Relu`, `Tanh`, `Sigmoid`, `Clip`, `Identity`, `Flatten`, `Reshape`, `Squeeze`, `Unsqueeze`, `Concat` and `Constant`. The game runs it with its own GDScript interpreter (no onnxruntime), so other ops are rejected.
- Input `obs`: `[1, 149]` float32, built by the game from the match snapshot (own and opponent state including shield meter, dizzy and air dodge, food and projectile slots, the two nearest sudden-death bombs, match timer and sudden-death phase, plus the fighter kits, the jab chain step and link window, the opponent hitstun and the combo dealt and taken). Indices 0-135 are the `bvs-schema-3` vector.
- Output `logits`: `[1, 21]`, the concatenated logits of eight discrete branches `[3, 3, 2, 5, 2, 2, 2, 2]` (`a0_move`, `a1_vertical`, `a2_jump`, `a3_attack`, `a4_special`, `a5_throw`, `a6_block`, `a7_dash`). The game takes the argmax of each branch, one decision every 4 physics frames. While holding food, `a5_throw` is exclusive (a punch and a throw never fire together), and in the air `a6_block` is the air dodge.
- `bvs.json` declares the schema version, observation size, action layout and `action_repeat`; a model whose sidecar does not match is refused with a visible reason.

## Fighter kits

Both Rafael (`kit_id` 0) and Matilde (`kit_id` 1) are played on both sides during training (random per
episode and per side, mirrors included); the observation carries each fighter's kit one-hot, so one
policy plays both kits and both match-ups.

## Evaluation

Played in the game with the same interpreter players use, full matches versus the scripted CPU, sides alternating.

| Opponent | Matches | Win rate |
|----------|---------|----------|
| cpu easy | 20 | 1.00 |
| cpu hard | 20 | 1.00 |
| cpu normal | 20 | 1.00 |

## Acceptance harness (combo and shield-spam gates)

172 matches (model vs model, model vs CPU, CPU vs CPU; both kits on both sides) against the acceptance
gates of `training/configs/default.toml` (sudden death length, food pickups, blocking, KO win rate,
shield-raise and shield-tap rate, combo length and duration, kit win-share balance, combo-damage balance).

| Gate | Verdict | Measured | Threshold |
|------|---------|----------|-----------|
| sudden_death_length | PASS | 2.54 | 45.00 |
| pickups | PASS | 6.64 | 1.00 |
| blocked_share | PASS | 0.23 | 0.70 |
| ko_win_rate_easy | PASS | 1.00 | 0.60 |
| shield_raises | PASS | 8.06 | 14.66 |
| shield_tap_share | PASS | 0.26 | 0.50 |
| max_combo | PASS | 5.00 | 6.00 |
| max_hurt_seconds | PASS | 1.35 | 1.50 |
| multi_hit_combos | PASS | 59.00 | 5.00 |
| kit_win_share | FAIL | 0.88 | 0.40 |
| combo_damage_gap | PASS | 0.09 | 0.20 |
| min_matches | PASS | 20.00 | 24.00 |

Overall (seed 1, kits `rm`): **FAIL**

## Files

- `model.onnx`: the policy (logits head only, opset 17).
- `bvs.json`: the schema sidecar (trained on commit `f7129e7fe2fb23320acd933b70cc21ff3a4ec05a`).

## License

mit