Five stages, running continuously
A self-driving car's software runs a continuous pipeline that mirrors the sense-think-act loop from general robotics, but broken into finer stages because the stakes and speed demand it. Sensing is raw data collection from cameras, lidar, radar, and other hardware. Perception turns that raw data into a structured understanding: this pixel cluster is a pedestrian, that point cloud is a parked truck, this lane marking curves left. Prediction then asks what each of those detected objects is likely to do in the next several seconds.
Planning takes the predicted futures of everything around the car and decides on a safe, comfortable trajectory for the vehicle itself, and control translates that trajectory into precise steering, throttle, and brake commands. All five stages run many times per second, and a failure or delay at any single stage can cascade into a bad outcome downstream, which is why each stage is engineered and tested largely independently before being integrated.