Case context: You are building a navigation system for an indoor delivery robot. You have 250,000 labeled images showing the locations of boxes, carts, and other hallway obstacles, but only 14,000 examples that map the robot’s raw sensor readings directly to steering commands.

Question: Given the data you have for intermediate steps in the task, what pipeline structure should you choose, and why?

Sample answer: I would choose a multi-stage pipeline. The large obstacle-location dataset is ideal for training an intermediate detector that finds objects in the hallway. Since there is much more data for that intermediate module than for the full sensor-to-command task, a staged design makes better use of the available labels than trying to force the whole problem into a single end-to-end model.

Key points:

• The obstacle-location labels are useful for an intermediate detection module
• A multi-stage pipeline fits a situation with plentiful intermediate data
• The detector can be trained directly on the large labeled set

Rubric: The learner must identify the obstacle-location data as training data for an intermediate module and conclude that a multi-stage pipeline is the better choice because it can exploit that larger dataset.