Tinyworldformer pairs a learned game engine with agents targeting the Raspberry Pi Pico 2 W. One network predicts the next game frame and reward from the current frame and action. The other chooses an action from the generated frame. For developers studying embedded inference, the project poses three separate questions: does the model fit in memory, how fast does it run, and does its generated world follow the game's rules?
The author reports that Pac-Man lite, the largest game, uses 600 KB of flash and 460 KB of SRAM. The README says Pico frame rate hasn't been measured. These are project-reported memory figures, not MrKeyoor hardware measurements. This is an explanation of the supplied implementation and results; the evidence does not establish a dated release event.
What the firmware implements
The supplied firmware at revision 93474fb7fb859e9dcebd6effb5373c0a243009a3 runs the world model and policy on core 1, while core 0 handles USB output. Its game loop selects an action from a human, a learning agent or a pretrained policy, then advances the learned environment. When learning is enabled, it updates the agent from that transition.
The source also formats an fps value for USB output, but that is not a published board benchmark. The calculation uses model and agent processing time. It excludes the human-play delay, and the loop can drop outgoing frames when the output channel is full. That value therefore cannot establish how many frames a user sees each second.
The world model supplies the environment; the policy chooses actions within it. This account uses the supplied source extracts. MrKeyoor did not build the firmware, execute an example, run a game, measure the board or verify its displayed output.
What the rule checks establish
The project's reported checks used the shipped agent for 5,000 generated steps per game at two model loops. Each loop reapplies the same transformer block. In the reported results, Pac-Man recorded 742 wrong game states and 38 wrong rewards. Flappy Bird recorded zero in both columns under that setup. These are the author's results for those trajectories, not proof of correctness across all policies or possible states.
The supplied Pac-Man evaluation source explains an important limit of the method. For each step, it reads state from the current generated frame, applies the real game rules to that state and the agent's action, then compares the expected result with the next generated state and reward. After a catch, the state comparison excludes random ghost respawn positions while checking Pac-Man, pellets and lives.
This checks whether each generated transition follows the rules from its current state. It does not establish that an entire generated episode matches a separate real-game episode from the same starting point. An earlier error can change the state from which later transitions are checked.
Developers investigating small learned environments can use the evaluation method to examine errors that an animation may hide. Anyone selecting the project for a specific Pico frame-rate requirement still needs board measurements. A useful next step is to measure processing time and displayed frame delivery separately, then repeat the rule checks with the intended agent.