Case context: A robotics team is training an aerial drone with reinforcement learning to carry packages to a charging bay and settle onto a marked platform. They define a reward function S(.) that gives a large penalty, for example S(T) = -800, when the drone collides with a wall, and a positive reward when it lands on the platform without crashing. But the current score only checks those two outcomes. It does not consider whether the final approach was jerky, whether the drone stopped in the center of the platform, or whether it disturbed nearby workers with unstable flight.

Question: What is the main weakness of this reward design, and how could it be improved so that it better captures a genuinely successful flight?

Sample answer: The main weakness is that the reward is too coarse. It treats every non-crash landing as equally good, even though some landings may be shaky, off-center, or uncomfortable for people nearby. A better design would add terms for approach smoothness, final position accuracy, and flight stability, then combine those terms into one score so the system is rewarded for the overall quality of the trajectory, not just the final outcome.

Key points:

• The current score only separates collision from non-collision landing
• A useful reward should account for smoothness, landing accuracy, and stability
• Reward design is hard because several objectives must be balanced
• The revised score should combine those objectives into a single measure

Rubric: Full credit: identifies that the reward ignores smoothness, landing precision, and stability, and proposes adding them to the score. Partial credit: identifies the weakness but does not describe an improvement, or describes an improvement without naming the missing factors. No credit: fails to identify the missing trade-offs.