GridWorld
Reinforcement learning in its purest form: paint walls and rewards on a grid, then watch value ripple outward from the goal one Bellman sweep at a time. Drag a wall mid-run and the whole value field — arrows and all — bends around your edit, live. Or freeze a cell and open the backup inspector to unfold the single number into its whole equation: four candidate actions, their slip-weighted outcomes, and the max that wins.
click or drag on the grid to paint
open the inspector to unfold how each cell's V(s) is computed from its four actions
edit anything mid-run — values reset and the sweep keeps going, so you can watch the field reroute live
slippery floor: 1 move in 5 slips sideways, so values sag next to the trap — the agent prices in its own clumsiness.
Bellman backups computed live in your browser — the numbers in each cell are V(s) from value iteration.