Pokemon Red, but the trainer is a neural network

What it does The project hooks the PyBoy Game Boy emulator into Stable Baselines 3 to train reinforcement-learning agents on Pokemon Red. A newer V2 script trains faster and with less memory, swapping a frame-based KNN novelty reward for a coordinate-based exploration incentive that pushes the agent to discover new map locations—reportedly as far as Cerulean. You can also load a pretrained model and step in interactively, or stream training sessions to a shared global map viewer.

The interesting bit The coordinate-based reward in V2 is the kind of practical hack that keeps RL from getting lost in visual noise; rather than comparing raw frames, it rewards the agent for visiting new (x, y) coordinates. The live broadcast wrapper turns solitary training into a multiplayer spectator sport, plotting every agent’s progress on a shared world map.

Key highlights

• V2 replaces pixel-similarity exploration rewards with coordinate-based progress, cutting memory use and training time.
• Includes a pretrained interactive mode where you can override the agent with keyboard input.
• Streams training sessions to a shared global map via a StreamWrapper integration.
• Ships with TensorBoard and optional Weights & Biases logging, plus static visualization tools.
• Has spawned follow-up work including an arXiv paper and the standalone PokeGym environment.

Caveats

• Setup is path-sensitive: the ROM must sit in the root directory while scripts run from subdirectories, and macOS or AMD GPU users need separate dependency workarounds.
• The README touts V2 as faster and leaner but offers no hard benchmarks; it only confirms the agent reliably reaches Cerulean.
• Python 3.10+ is recommended, with other versions explicitly untested.

Verdict Anyone curious about applied RL on classic games—or looking for a pre-built PyBoy wrapper with logging and broadcasting—should dig in. If you need a general-purpose Game Boy RL framework rather than a Pokemon-specific demo, check the spin-off PokeGym instead.