FootsiesGym: A Fighting Game Benchmark for Two-Player Zero-Sum Imperfect-Information Games

2026-07-07 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Gaming & Interactive Media · Depth: Expert, quick

Summary

FootsiesGym, an open-source environment released on 2026-07-07, facilitates reinforcement learning research in two-player, zero-sum, imperfect-information games. Built upon HiFight's minimalist 2D fighting game Footsies, it specifically isolates the cyclic, non-transitive strategic interactions of fighting game "neutral play" for efficient analysis. The environment features a vectorized simulator, enabling high-throughput training on standard hardware, which enhances accessibility and reproducibility for researchers. Its design is detailed, and several reinforcement learning algorithms have been benchmarked within it. The authors also discuss various open research directions that FootsiesGym enables, with the code publicly available at https://github.com/como-research/FootsiesGym.

Key takeaway

For Machine Learning Engineers developing agents for complex competitive environments, FootsiesGym offers a robust, open-source benchmark to accelerate research. You can leverage its vectorized simulator for high-throughput training of reinforcement learning algorithms on standard hardware, significantly reducing iteration times. This environment is ideal for exploring strategies in two-player, zero-sum, imperfect-information games, particularly focusing on "neutral play" dynamics. Consider integrating FootsiesGym to validate new RL approaches and contribute to fighting game AI research.

Key insights

FootsiesGym provides an accessible, high-throughput benchmark for RL in complex two-player, zero-sum, imperfect-information fighting games.

Principles

• Isolate core strategic interactions.
• Vectorized simulation boosts training throughput.
• Open-source environments enhance reproducibility.

Method

The environment design involves building on a minimalist game, isolating specific strategic interactions, and implementing a vectorized simulator for efficient algorithm benchmarking.

In practice

• Benchmark new reinforcement learning algorithms.
• Analyze fighting game "neutral play" strategies.
• Develop agents for imperfect-information games.

Topics

• Reinforcement Learning
• Fighting Game AI
• Zero-Sum Games
• Imperfect Information Games
• Game Theory
• Open-Source Software

Code references

• como-research/FootsiesGym

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Bench-MFG: A Benchmark Suite for Learning in Stationary Mean Field Games — Bench-MFG standardizes Mean Field Game evaluation through a diverse benchmark suite and procedural generation.
• ShopGym: An Integrated Framework for Realistic Simulation and Scalable Benchmarking of E-Commerce Web Agents — ShopGym offers a scalable framework for creating realistic, controllable, and reproducible e-commerce web agent evaluation environments.
• In game theory, generalists sometimes win out over specialists — General-purpose policy gradient methods can surprisingly outperform specialized game-theoretic algorithms in imperfect-information games.
• Mahjax: A GPU-Accelerated Mahjong Simulator for Reinforcement Learning in JAX — Mahjax provides a high-throughput, GPU-accelerated JAX environment for *tabula rasa* reinforcement learning in complex imperfect-information games.
• Chess-World-Model: A 10M-Game Benchmark for Exact State Tracking from Chess Move Sequences — Large-scale, out-of-distribution benchmarks reveal state-tracking failures in models that scale alone conceals.
• [P] Vibecoded on a home PC: building a ~2700 Elo browser-playable neural chess engine with a Karpathy-inspired AI-assisted research loop — An AI-assisted research loop can rapidly develop high-performance neural chess engines on consumer hardware.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.