On July 8, Chase McDonald, Nathan Tsang, and Wesley N. Kerr published an arXiv preprint detailing FootsiesGym, a reinforcement learning environment based on a minimalist 2D fighting game [S1][P2]. It focuses on a category of AI challenges that has historically been difficult: two-player, zero-sum games with hidden information. The core inquiry is if a simplified fighting game can provide a clean, rapid testing ground for complex strategic interactions in machine learning, with implications extending far beyond video games.

Why fighting games break AI

AI benchmarks typically fall into two categories. Perfect-information games, such as chess and Go, reveal all pieces and rely on search depth. Imperfect-information games, like poker, involve hidden cards and require bluffing and reasoning under uncertainty. While the former is mostly solved, the latter remains difficult, as current testbeds are often either overly simplistic, like rock-paper-scissors, or too complex, like full commercial games with numerous mechanics [S1].

Fighting games occupy a challenging middle ground. Two characters battle at close range, with options to attack, block, or move. Neither player knows the opponent's action until it is too late to react. The interactions are cyclic—attack beats throw, throw beats block, and block beats attack—and non-transitive, meaning no single strategy dominates [S1]. This is known as the "neutral" phase in fighting games, where neither player has an advantage and both seek an opening. This type of strategic loop is exactly what causes AI agents to struggle in real-world adversarial scenarios.

FootsiesGym extracts this neutral phase and removes all other elements like health bars, combos, and stages. The result is a strategic core that is simple enough for efficient analysis yet complex enough to avoid trivial solutions [S1].

A vectorised simulator on standard hardware

The environment includes a vectorised simulator, allowing thousands of game instances to run in parallel on standard hardware rather than a dedicated GPU cluster [S1]. For RL researchers, this means the difference between overnight experiments and week-long runs. The codebase is open-source under GPL v3.0, available on GitHub, and primarily written in Python, with C# components from the original Unity game [P3][P4]. The repository was established in March 2025, providing over a year of quiet development prior to this preprint [P3].

The authors test multiple RL algorithms within the environment and explore the research directions it opens up [S1]. The preprint's abstract does not include specific win rates or numerical scores, as the benchmark is intended as community infrastructure rather than a leaderboard claim [S1].

What it means

The primary contribution of FootsiesGym is not the game itself, but the problem structure it presents. The combination of imperfect information, zero-sum dynamics, two-player design, cyclic interactions, and non-transitivity describes a wide array of real-world strategic situations, from auctions to cybersecurity and negotiations. Previously, researchers had to choose between overly simplistic toy problems and overly complex real games. FootsiesGym offers a middle ground: a game learnable in five minutes where the optimal strategy is neither obvious nor memorisable.

For those unfamiliar with the field: imagine training an AI for rock-paper-scissors, but with costs, timing, and punishable counters. This is what FootsiesGym simulates. The AI cannot just memorise a winning move; it must learn to adapt by modelling the opponent's likely actions and exploiting patterns while remaining unpredictable.

FootsiesGym poses a sharper question: can an agent outthink an opponent that is actively trying to outthink it?

What it means for business

The primary audience is academic and applied AI labs, not commercial game studios, but the downstream effects impact several operator types:

  • Two-person AI consultancies focused on adversarial ML, such as anomaly detection and red-team simulations, gain a free, fast testbed for multi-agent strategies that previously required custom environments.
  • Game studios developing AI opponents can use FootsiesGym as a research scaffold. Its minimalist design clarifies why an agent behaves in certain ways, which is crucial for tuning difficulty and avoiding exploitable patterns.
  • Cybersecurity firms testing adversarial agents receive a benchmark where the strategic structure of hidden information and cyclic counters closely maps to attacker-defender dynamics.

The practical change for researchers this quarter: workflows that once required licensing a commercial engine or building a custom simulator can now begin with a git clone and a Python environment [P3]. The vectorised simulator allows for iteration cycles in hours rather than days [S1].

What we don't know yet

The preprint has not undergone peer review [S1]. All claims regarding capabilities and design are attributed to the authors, and the environment's performance, stability, and accessibility on other machines remain unverified. No third-party researchers have independently validated the benchmark, and the paper lacks specific numerical results for the tested algorithms [S1].

The public GitHub repository had only three stars and one open issue at the time of research [P3], indicating early-stage adoption. Whether the broader RL community adopts FootsiesGym as a standard benchmark or treats it as an internal tool will depend on independent reproduction and community-contributed baselines.

A related project, vlongle/Imperfecto, provides a Python library for imperfect-information games with self-play and exploitative algorithms under an MIT licence [P5], serving as a potential complement or competitor, though the projects differ in scope and design.

The next concrete signal to watch is whether the paper is accepted at a peer-reviewed venue like NeurIPS or ICML, and whether independent labs publish baseline results. Until then, it remains a promising tool with an open invitation and an open question.

If this deep-dive into AI research infrastructure is useful, subscribe to keep reading—we will be tracking FootsiesGym and the broader benchmark landscape as it evolves.

Sources

Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.