Wombo Combo: cooperative Fox policies for Melee, with the Frisson Labs wave logo.

Wombo Combo

Wombo Combo is an experiment in Super Smash Bros. Melee that we made for fun. Two Fox bots team up on a dummy opponent to see how far they can take a combo.

Online play: Please do not use these models in online matches without explicit permission from the other participants and the relevant platform or event organizers. Follow the platform's rules.

Inference project · Exact input schema · Reyn collection

Let's see some combos

Wombo Combo: two Fox bots keep a combo going on a dummy opponent, ending in a KO.

A highlight from the blog, shown at normal speed.

The two Fox bots play on Final Destination against a dummy Fox that doesn't fight back. Damage starts at 0% and works normally, with four stocks and friendly fire off. The bots choose their actions as they play.

We picked this clip as a highlight; results vary between games. We haven't established how well this model plays with human partners, against human opponents, on other stages, with friendly fire on or in general 2v2 play. The dummy clip also doesn't tell us how it would do against an opponent that fights back.

Inference

Install the minimal PyTorch inference package. Download base.pt, adapter.pt, config.json and feature_schema.json from this repository into combo-files/. Checksums are included in checksums.sha256.

from reyn import Combo

model = Combo.load("combo-files")
inputs = model.example_input()
frames, features = inputs
out = model.predict(
    frames,
    features,
)
print(out.action)
labels = out.labels.as_tensor()
logits = out.logits
model.reset(seed=0)

Each bot's row needs the base observation and 186 partner/scoring-state features in the documented order. Replace the synthetic values with consistent model inputs. The bots share weights and keep separate prediction histories; you can reset each row independently. The runtime uses the saved, fixed partner-action table and checks the exact base and adapter identities. Wombo Combo takes no language phrase.

Frisson Labs · Based on Faynt.

License

Frisson Labs' inference code and model weights are released under the MIT License. See the third-party notices for upstream attributions.

The online-play warning is guidance; the MIT license terms are unchanged.

Downloads last month
50
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Collection including frisson-labs/Wombo-Combo