A new open-source benchmark called FootsiesGym puts reinforcement learning agents head-to-head in a stripped-down fighting game designed to isolate the hardest strategic problems in the genre.
Built on HiFight's minimalist 2D title Footsies, FootsiesGym is a two-player, zero-sum environment where neither player has full information about the other's intentions — the same basic structure as poker, but with reflexive, cyclic move interactions instead of cards. The researchers behind it provide a vectorized simulator tuned for high-throughput training on standard hardware, meaning you don't need a server farm to run experiments. Several reinforcement learning algorithms are benchmarked out of the box, and the code is public on GitHub.
Why does a tiny fighting game matter to AI research? Most game benchmarks lean on perfect-information environments like chess or Go, or sprawling ones like StarCraft II that are expensive to run and hard to isolate variables in. FootsiesGym carves out a specific, underexplored niche: the non-transitive "rock-paper-scissors" dynamics of fighting game neutral — the phase before either player commits to an attack — where no single strategy dominates and agents must model an opponent who is also adapting. That's a closer analog to real-world adversarial settings than a chessboard.
The benchmark arrives as multiagent and adversarial RL research is growing faster than the tooling around it. Whether FootsiesGym becomes a standard fixture or a footnote depends on whether the wider community finds the Footsies abstraction tight enough to generalize — fighting game "neutral" is famously hard to define even for human players.