Essay

Explain why a navigation stack can fail even when its parts look strong.

Question: Suppose a campus delivery robot uses three modules: people detection, cart detection, and route selection. Each module performs close to human level when judged on its own input, but the whole system still chooses poor paths. Why can this happen, and how can it be improved without making the route selector analyze camera frames directly? Describe the purpose of the extra detector.

Sample answer: The weakness comes from missing context in the overall system. Even if the detector modules are accurate, the route selector may still lack scene details that matter for safe movement, such as aisle boundaries or blocked passageways. One way to fix this is to add an aisle-boundary detector before the route selector. That detector reads the camera image, extracts structured boundary coordinates, and sends those coordinates to the route selector. The route selector then gets the needed scene layout in a compact form, instead of being forced to process raw pixels.

Key points:

  • The system performs poorly because the final decision module does not receive all the information it needs.
  • Good performance of separate modules does not guarantee good end-to-end performance.
  • An added detector can supply the missing layout information.
  • The route selector should receive structured outputs, not raw camera images.

Rubric: The response must explain that the problem is a system-level lack of needed information, not simply weak individual detectors. It must identify an added detector for the missing scene feature and state that its structured output is passed to the route selector instead of sending raw images to that module.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI

Related