Question: Why can a direct end-to-end model for controlling a delivery drone be hard to train compared with problems where labeled input-output examples are easy to collect?

Sample answer: A direct end-to-end system for a delivery drone would need a very large set of matched examples such as camera frames paired with the control actions taken at that moment. Those labels are costly to obtain because they do not already exist in a convenient dataset. The team would need to equip many drones with logging hardware, fly them repeatedly in different weather, lighting, and obstacle conditions, and record enough varied flights to cover rare situations. That makes data collection slow and expensive.

Key points:

• Needs many paired image-and-control examples.
• Those examples are expensive and time-consuming to gather.
• The vehicles need extra instrumentation for logging.
• Many flights are needed to span diverse conditions.

Rubric: 1 point for identifying the need for matched image-and-control pairs. 1 point for saying the data is costly and slow to collect. 1 point for mentioning special logging or instrumentation on the drones. 1 point for noting that extensive flights are required to capture diverse situations.