Case context: A city logistics startup is building an autonomous delivery cart. One team proposes a single neural network that reads street-camera video and directly outputs steering and speed commands. Another team proposes a staged system that first identifies lanes, obstacles, and pedestrians, then computes the route. The company has only a small set of complete driving traces but a large labeled set of object and lane annotations.

Question: Which design is the better choice right now, and why?

Sample answer: The staged modular design is better at this stage. A fully direct model is more appealing when the team has enough complete driving examples, but this company has few full traces and many labeled intermediate examples. Because their data is much richer for the perception steps than for direct control, a modular pipeline fits what they can train reliably today.

Key points:

• Recommend the modular or staged design.
• Note that complete driving traces are limited.
• Explain that the available labels support intermediate perception tasks better than direct control.

Rubric: The student must recommend the modular approach and justify it by explaining that the available data favors intermediate perception modules over direct control.