Wayfinder, an RL Observatory

Field notes

Wayfinder models navigation as a visible grid-world experiment rather than hiding training behind an opaque demo. Its controls expose both model-free choices, such as Q-learning and SARSA, and Dyna-Q planning, alongside parameters for exploration and discounting. Visitors can inspect individual cell values, compare a learned path with chance, and watch a 3D simulation. The supplied page begins with no completed episodes or environment steps, so the training results are not yet present in the captured state.