State machines and hand-tuned controllers
Classical robot programming, the kind you'd use to build a pick-and-place arm on a factory line, typically works through explicit state machines. An engineer defines a fixed sequence: move to position A, close the gripper, check a force sensor to confirm a grasp, move to position B, open the gripper. Each transition is hand-coded, each controller gain is hand-tuned, and each edge case, like the object being slightly off-center, needs to be anticipated and explicitly handled. This approach is precise and predictable when the environment is well-controlled, which is exactly why it has dominated industrial robotics for decades.
The weakness shows up the moment the environment varies in a way nobody anticipated. If the object is a slightly different shape, if lighting changes how a vision system detects an edge, or if the task itself changes from 'stack boxes' to 'fold laundry,' the hand-coded state machine typically breaks or needs to be rewritten from scratch. Generalization was never the goal of classical control, reliability within a narrow, specified envelope was.