Description
Learning Pokémon With Reinforcement Learning is a research project that trains a reinforcement learning agent to beat the 1996 game Pokémon Red. The authors say they have been developing it since 2020, and that as of February 2025 they could beat the game with a policy under 10 million parameters and minimal simplifications. The site stresses that the output is a technique for producing solutions to Pokémon, not a policy that can play the game in general. It is organized as chapters: an introduction to RL, the environment with its observations, rewards and policy, building and running the system, including reading game RAM and tracking metrics, and concluding thoughts on results and future work. An appendix explains Pokémon battling, the eight badges and field exploration. The code is open source and linked from the site. The page has a changelog that starts with a March 2025 release, and the authors invite contact through a Discord server or by email.…
Highlights
- Reinforcement learning agent for Pokémon Red
- Chapters on environment, rewards and policy
- Open source code
- Metrics and visualization of training
Ratings & Reviews
No ratings yet
Be the first to rate Pokémon RL.
People also visit
Jev Plays Pokémon Red
Jev, an AI decision model, plays Pokémon Red all the way to the Hall of Fame

stable-baselines3
Trains and evaluates single-agent reinforcement learning with Stable Baselines3
Engine
A minimal, complete, Pokémon battle simulation engine optimized for performance
Pokémon RL FAQ
What is this project?
A reinforcement learning project that beat Pokémon Red with a policy under 10 million parameters.
Is the output a playable agent?
The authors say it is a technique for producing solutions to Pokémon, not a policy that can play the game in general.
Where is the code?
On GitHub under drubinstein/pokemonred_puffer, linked from the site.
How do I contact the authors?
Through the Discord server or by email, both linked from the site.







