perception–planning–control architecture
A perception–planning–control architecture is the standard way of organizing a robot's software into three stacked layers, like an assembly line for turning raw sensor data into motor commands. Perception is the first layer: it takes the flood of numbers coming from cameras, laser scanners, and other sensors and turns them into a tidy understanding of the world — where the walls are, where the cup is, how fast it is moving. Planning is the middle layer: given that understanding and a goal, it decides what to do — which path to follow, which object to grab, in what order. Control is the last link in the chain before the motors: it takes the plan and produces the exact, fast, moment-to-moment commands that drive the wheels or joints so the robot actually carries it out smoothly.
The reason engineers split the software this way is that each layer has a very different rhythm and job, so it is cleaner to build and fix them separately. Perception may chew on a whole camera image and only update a few times a second; planning might think for a fraction of a second about a route; control runs hundreds or thousands of times a second to keep a joint steady against gravity and friction. Information mostly flows forward — sensors feed perception, perception feeds planning, planning feeds control — and you can swap out or improve one layer (say, a smarter object detector) without rewriting the others, as long as they agree on what to hand across the boundary.
In practice the picture is rarely a strict one-way street. Fast feedback often loops back: control reports whether a motion succeeded, perception keeps watching to catch surprises, and planning replans when the world does not match expectations. Many modern systems also blend the layers — for example, a learned model that maps images almost directly to actions, collapsing perception, planning, and control together. Even then, this three-part way of thinking remains the mental map most roboticists use to reason about where a behavior is decided and where a bug might live.
In a home robot fetching a mug, perception spots the mug on the table, planning chooses a collision-free arm path, and control feeds the joint motors the precise currents that trace that path.
Three layers, three jobs: understand the scene, decide a path, then move the motors to follow it.
This is the software-layering view of the same idea behind the sense–plan–act loop; the names emphasize the three software stages rather than the repeating cycle.