World model runs and plays games on Raspberry Pi Pico 2W
Original: A tiny world model running games on a Raspberry Pi Pico 2 W
Why This Matters
Shows ML inference and RL training are reaching sub-$10 embedded hardware.
Developer SheepSeb built two small neural networks that generate and play games entirely on a Raspberry Pi Pico 2W microcontroller, using looped transformers to keep memory usage low enough for embedded hardware.
A GitHub project called tinyworldformer demonstrates that world models — neural networks that predict future game states — can run on cheap microcontrollers. The setup uses two networks: one generates the next frame and reward given a current frame and action (the world model), while a second network picks actions based on those generated frames (the actor). Both run on a Raspberry Pi Pico 2W, a board that costs a few dollars and includes a floating-point unit.
The key trick is looped transformers: rather than stacking multiple distinct transformer layers, a single block is applied repeatedly with shared weights. This cuts memory footprint sharply without entirely sacrificing model capacity. The developer wrote the target games in NumPy first, which allowed frame-by-frame verification of what the world model was generating versus ground truth.
A streaming reinforcement learning agent was later added, letting the board learn to play inside its own generated dream environment — no external compute required. The project is written in Rust and Python, with quantization applied to make inference viable on constrained hardware. It has attracted 159 stars on GitHub since posting.