Behavior-cloned Super Smash Bros. Melee bots trained on human Slippi
replays. Eight character-specific ~20M-parameter transformers that take
a 180-frame window of game state and output controller inputs (main
stick, c-stick, shoulder, buttons) at 60 Hz. Each model plays over
Slippi Online Direct Connect through Dolphin + libmelee.
Training data:
erickfm/melee-ranked-replays
— ranked Slippi replays (master/diamond/platinum tier) per character.
Base architecture: Shaw-relative-position causal transformer
(d_model=512, 6 layers, 8 heads, seq_len=180). Bootstrapped from
HAL (Eric Gu) and since diverged.
Defining MIMIC changes over HAL: 7-class button head with a
distinct TRIG class for airdodge/wavedash (HAL's 5-class head can't
represent airdodge and thus can't wavedash); v2 shard alignment that
fixes a subtle post-frame-gamestate leak in the training targets
(see research-notes-2026-04-11c); the digital-L-press fix in
decode_and_press (research notes 2026-04-13) without which no
7-class BC bot wavedashes.
Current checkpoints (retrained on 2026-04-20 baseline)
Retrained on the post-schema-drop (13 numeric cols), new-transforms
(tanh_scale / linear_max / log_max for velocity / hitlag /
hitstun) basis. See research-notes-2026-04-20.md in the MIMIC repo
for methodology + results analysis.
Character
Run
Train games
Val loss
Step
Fox
fox-20260420-baseline
31,030
0.7144
32768
Falco
falco-20260420-baseline
20,882
0.7487
31392
Marth
marth-20260420-baseline
11,759
0.6664
31065
Sheik
sheik-20260420-baseline
51,751
0.6566
26160
Captain Falcon
cptfalcon-20260420-baseline
17,557
0.7368
watchdog
Luigi
luigi-20260420-baseline
2,290
0.7460
watchdog
Peach, Jigglypuff, and Ice Climbers remain on pre-2026-04-20
schemas:
peach-20260420-baseline (val 0.6322) was trained on the 22-col
schema before the schema drop — loadable via its pickled config.
puff and ice_climbers missed the 2026-04-20 retrain cycle due
to a download-script bug; their existing HF checkpoints are on the
old schema. These two are incompatible with the current 13-col
inference code path. Will be retrained in a follow-on cycle.
Each character directory is self-contained — the JSONs are the exact
metadata used during training, copied verbatim from the data dir so
any inference script can load them without touching the MIMIC repo.
HAL's original 5-class head (A / B / Jump / Z / None) has no TRIG class
and structurally can't execute airdodge, which means HAL-lineage bots
can't wavedash. MIMIC's 7-class encoding plus a fix for
decode_and_press (which was silently dropping the digital L press
until 2026-04-13) is what enables the wavedashing in the replays.
5 velocities (scale=5 for self, scale=10 for attack)
linear_max
x / max
hitlag_left (max=20)
log_max
log1p(clamp(x,0,max)) / log1p(max)
hitstun_left (max=120)
Plus categorical embeddings: stage(4d), 2× character(12d),
2× action(32d). Plus the previous-frame controller state as a 56-dim
one-hot (37 stick + 9 c-stick + 7 button + 3 shoulder).
Total input per frame: 184 dimensions → projected to 512.
Earlier builds (pre-2026-04-20) used a 22-col numeric schema that
included invuln_left and 8 ECB corners. Those columns turned out to
be structurally zero for our .slp parse path — libmelee never
populates them — so they were dropped from the schema. See research
notes 2026-04-20 for the audit. Checkpoints trained pre-drop
(peach-20260420-baseline) still load via their own pickled config
but use the 202-dim projection path.
Training
Model preset: mimic (20M params)
Optimizer: AdamW, LR 3e-4, weight decay 0.01, no warmup
Mixed precision: BF16 AMP with FP32 upcast for relpos attention
(prevents BF16 overflow in the manual Q@Kᵀ + S_rel computation)
Max samples: 16.78M (≈ 32,768 steps at eff-batch 512)
Watchdog: patience=12 evals on val-plateau — some chars finish early
Reaction delay: 0. v2 shards have target[i] = buttons[i+1], so
rd=0 matches inference — do NOT use --reaction-delay 1 or
--controller-offset with v2 shards.
--self-inputs is required even on v2 shards. Runs without it
drop the controller-history input entirely and land at val loss ~2.3.
Typical wall-clock per char on 2×RTX 5090: 10-15 min download/extract
20 min parallel norm_stats bootstrap + 45-120 min sharding
(depending on char, cptfalcon and sheik are the longest) + ~50 min
training = 2-4 hours.
Known limitations
Character-locked. Each model only plays the character it was
trained on. No matchup generalization. Multi-character training
with a character embedding is a natural next step but not done.
Small-dataset overfitting on Luigi / Ice Climbers. Luigi has
~2K training games; IC around 5K. Their _bestloss.pt is
early-stopped — either by the patience=12 watchdog during this
cycle or by inspection in prior cycles. Play quality varies.
Edge guarding and recovery weaknesses. Bots don't consistently
go for off-stage edge guards or execute high-skill recovery
mixups. The training data has these in it, but BC bots under-sample
long-tail strategic decisions.
No Matchmaking / Ranked. The Discord bot only joins explicit
Direct Connect lobbies. Do NOT adapt it for Slippi Online Unranked
or Ranked — libmelee's README explicitly forbids bots on those
ladders, and Slippi has not yet opened a "bot account" opt-in
system.
Acknowledgments
Eric Gu for HAL, the
reference implementation MIMIC is based on. HAL's architecture,
tokenization, and training pipeline are the foundation.
Vlad Firoiu and collaborators for
libmelee, the Python interface
to Dolphin + Slippi.
Project Slippi for the Slippi Dolphin fork, replay format, and
Direct Connect rollback netplay. https://slippi.gg
