Case context: You are maintaining an automated driving stack for an urban robotaxi. The system includes a traffic-light recognizer, a bicycle detector, and a steering-and-speed control module. You want to diagnose weaknesses and raise overall quality, but you are not using a formal evaluation framework.

Question: How can you use a simple part-by-part error review to choose which module deserves the first improvement effort? What comparisons should you make?

Sample answer: A practical way to do this is to compare each subtask with the level reached by a strong human on the same subtask. I would check the traffic-light recognizer against a human judge of traffic lights, the bicycle detector against a human’s bicycle detection ability, and the full driving stack against the performance of an experienced human driver. The module with the largest gap to the human benchmark is the one to attack first, because it is likely limiting the end-to-end result the most.

Key points:

• Compare the traffic-light recognizer with human performance on traffic-light identification.
• Compare the bicycle detector with human performance on bicycle detection.
• Compare the complete driving system with human driving performance.
• Use the largest gap to decide where the next engineering effort should go.

Rubric: The answer must name the specific comparisons to make for each component and for the full system, and it must explain that these comparisons are used to find the weakest part to improve first.