Reward-Conditioned Attention: How Reward Design Shapes What Autonomous Driving Agents See

Researchers have demonstrated that the reward structures used to train autonomous driving agents significantly alter how those systems distribute their internal attention while navigating. By testing models with identical architectures but different reward goals, the study shows that subtle changes in feedback mechanisms fundamentally reshape which environmental features a vehicle prioritizes. These findings suggest that reward engineering is a critical factor in determining the visual decision-making processes and potential safety outcomes of self-driving artificial intelligence.