A tiny world model running games on a Raspberry Pi Pico 2 W (github.com)

A pioneering project has successfully developed a compact world model capable of generating and playing games on a Raspberry Pi Pico 2 W using two small neural networks. The first network, the world model, predicts the next frame and reward based on the current frame and the player's action, while the second, the actor, decides the action to take from the generated frames. This innovative architecture utilizes looped transformers, allowing repeated application of a single transformer block to conserve memory, critical for running on the low-resource Pico 2 W. This achievement is significant for the AI/ML community as it demonstrates the feasibility of deploying sophisticated AI models on microcontrollers, expanding the potential for real-time inference in edge computing. The project features various game models, optimizing parameter counts and computational costs for each game. For instance, the Breakout model operates with only 26,000 parameters and achieves over 99% accuracy in frame prediction. Notably, the introduction of a streaming reinforcement learning agent enables the model to learn and adapt as it plays, illustrating a step forward in the capabilities of embedded AI systems and their applications in interactive environments.