Paul Merceur

four rooms RL

real-time PufferLib demo running in the browser. i built the Four Rooms environment in C so the same env can train fast locally and compile down to WASM for the site.

the agent is a tiny PufferNet policy exported from a local checkpoint: 7x7x3 observations, two 128-wide hidden layers, and about 118k float32 parameters.

click on the demo, hold Shift and press W/A/D or the arrow keys to take control; the agent takes over when you release Shift.

environment based on Minigrid Four Rooms, code available on GitHub.