Subsumption and Layered Behaviors
Instead of one big plan, a behavior-based robot runs several simple behaviors at once. A fixed priority decides which one drives the wheels right now. Rodney Brooks called this the subsumption architecture, and the first Roomba worked this way.
How the layers combine
Each behavior watches the sensors and can ask to control the wheels. A higher layer suppresses every layer below it while it is active. Here, avoid wins whenever an obstacle is close. Otherwise seek steers toward the light. If seek is turned off, wander drives in slow random curves.
What to notice
Turn off avoid and the robot heads straight for the light, crashing into anything in the way. Turn off seek and it explores the arena safely but aimlessly. With all three on, smart-looking behavior appears even though no layer knows about the others. When the robot reaches the light, the light jumps to a new spot.
Layered behaviors are fast and robust, but hard to scale to complex tasks, because it gets hard to predict how many layers will interact. Modern robots often keep a reactive safety layer like avoid, and put a planner or a behavior tree above it.