● Covered by 1 source · 1 reportMedium impact1 neutral
Tiny World Model Runs Games on Raspberry Pi Pico 2 W Using Looped Transformers
Key points
- Two neural networks generate and play games simultaneously.
- Looped transformers reduce memory usage for microcontrollers.
- Runs on Raspberry Pi Pico 2 W, utilizing its FPU.
- Actor network learns to play within the generated game environment.
Tiny World Model Architecture
The system employs two small neural networks to operate games on a Raspberry Pi Pico 2 W. One network generates the game by predicting the next frame and reward based on the current frame and action. The second network, an actor, plays the generated game by selecting actions based on the predicted frames.
Looped Transformers for Memory Efficiency
The project utilizes looped transformers, applying a single transformer block multiple times with the same weights instead of stacking different layers. This design choice significantly reduces memory consumption, making it feasible to run complex models on microcontrollers like the Pico 2 W, which was selected for its cost-effectiveness and integrated FPU.
On-Chip Operation and Learning
Both the world model and actor networks run on core 1 of the RP2350 chip, while core 0 manages USB and display functions. The actor network operates solely on frames generated by the world model, and the world model receives actions only from the actor. The system supports a streaming RL agent that can learn to play games from scratch directly on the board, allowing for dynamic adaptation within the generated environment.
Game Implementation and State Representation
Each game is implemented in a few hundred lines of NumPy code. Games render to a small grid of token IDs, with the entire game state contained within a single frame. This includes hidden states like a ball's heading or a minimap in games like Doom-lite, which are encoded directly into the frame's tokens.
✨ This summary was generated by AI from the outlets' reporting listed below. It is not independently verified and may contain errors — check the original sources. How BrevFeed works →
The daily brief
One email each morning: the day's tech stories, clustered across outlets and summarized. No account needed.
One email a day. Unsubscribe in one click, any time.
Today's brief
Spend a few minutes, get the whole day. Every topic's top stories in one hands-free rundown — listen, watch, or read the transcript.
▶ Play today's brief
New every morning, and the back catalogue is archived by date.
Primary sources
GitHub cmc_internal/api GitHub github/collect GitHub SheepSeb/tinyworldformer GitHub _private/browser GitHub get-started/accessibility GitHub en/copilotReporting from
Researchers developed a tiny world model using two small neural networks and looped transformers to generate and play games on a Raspberry Pi Pico 2 W. This approach allows a microcontroller to learn and play games by predicting future frames and actions, demonstrating efficient AI model deployment on resource-constrained hardware.