pyboy

Playing Pokemon Red using TypeSafe Jev

Autonomous Pokémon Red agent using local Qwen3-VL and PyBoy. Zero API calls, zero cost, pure local VLM reasoning and computer vision.

Gameboy (Color) Environments in Gymnasium.

Super Mario Land Reinforcement Learning with Pufferlib

Local Pokemon Red LLM-agent harness with live gameplay UI, structured traces, save states, replay, and turn-by-turn agent observability.

Work in progress: agents playing the real Pokémon Blue ROM on PyBoy, with every badge, catch and evolution verified in cartridge RAM.

PPO reinforcement-learning agents that learn to play classic games from scratch - DOOM, Pokemon Red, and Street Fighter II - one standalone trainer per branch. Experimental R&D; bring your own ROMs.

Explores reinforcement learning and its core algorithms through a PyBoy-emulated instance of Pokémon Blue.

A learning experiment exploring how to teach an AI model how to play the game Tetris for GameBoy using StableBaselines3 and PyBoy.

Hold-aware DQN training for Pandora's Blocks through PyBoy

Twitch-plays-Pokemon style bot for Slack using PyBoy

AI Gym environment for Castlevania: The Adventure (Game Boy), using the PyBoy emulator.

Completed reproducible Pokémon Red study: 8.24M self-generated actions reached Route 1, but frozen evaluation found zero durable skills.

PyBoyとGBDK(GameBoy Developers Kit)の例

ROM-free OpenRappter agent for a local Copilot-powered Pokemon Red playthrough with PyBoy

🎮🌟A Reinforcement Learning (RL) project using Double Deep Q-Network (DDQN) to train an agent to play Kirby's Dreamland.

An AI agent for Street Fighter that uses reinforcement learning to improve through repeated matches, adapting its actions and strategies based on gameplay experience.

Agente de Reinforcement Learning (PPO) treinado para jogar Zelda: Link's Awakening no Game Boy via emulador PyBoy, com observacao multimodal.

Canonical RAPP cartridge for a local Copilot-powered Pokemon Red playthrough

A fixed agent harness for Pokémon Red where the model is the only variable: same tools, memory, prompt and observation format across 5 models from 4 vendors, turn-matched and eval-integrity gated. Live leaderboard with per-seed spreads and every exclusion.