RLCard is a toolkit for reinforcement learning research in card games, providing ready-made environments and baseline agents for Blackjack, Leduc Hold’em, Texas Hold’em, Dou Dizhu, Mahjong, UNO and more. Card games are a classic testbed for imperfect-information AI — players must reason about hidden state and opponents — and RLCard standardises them the way Gym standardised control tasks. The project, from Texas A&M’s Data Mining Lab, has around 3,600 stars on GitHub and is MIT licensed.
Each game ships as an environment with a consistent API: reset, step, legal-action masks and configurable state representations, plus rule-based and random agents for sanity checks. Built-in algorithms include DQN, NFSP and Deep CFR, so researchers can reproduce known results and plug in their own agents for comparison. Dou Dizhu and Mahjong — large, culturally significant games with huge state spaces — are first-class citizens rather than afterthoughts, and the toolkit includes trained models and an online demo. RLCard has been used in coursework and in research on multi-agent learning, game solving and opponent modelling.
RLCard is completely free and open source under the MIT license. Training agents only costs your own compute time.
Reinforcement learning researchers, students and game-AI hobbyists who want a standard benchmark for imperfect-information games without implementing the rules and environments from scratch.