Safety Gate
Let a rover move only inside its safe set.
How to play
Guide the rover from the bottom left to the dock at the top right in twelve control ticks. Its learned policy proposes a direction. Allow a safe proposal, project an unsafe one onto a safe route, or brake and spend a tick. Projecting a proposal that was already safe costs an extra tick. Red cells are obstacles; the yellow outline marks a safe proposed destination.
The systems concept
A learned policy can propose an action, but a separate safety gate decides what crosses into physical actuation. The toy grid shows the distinction between accepting a safe proposal, projecting an unsafe one to a safe alternative, and braking when movement must stop. Real systems need state estimation and a verified stopping model to make such decisions.
Explore Volume IV: Physical AI or return to the playground.