Tsinghua AIR team reveals the fundamental differences between human and autonomous driving algorithm visual attention.

In February 2026, the AI team of the Tsinghua University Institute of Intelligent Industry Research published a study titled "Human vs. Algorithm Visual Attention in Driving Tasks" in "npj Artificial Intelligence." Using autonomous driving as the carrier in this safety-critical field, for the first time through a dual-track design of "human eye tracking experiments + algorithm comparative verification," the essential differences between human and algorithm visual attention were systematically dissected. The core value lies in proposing a three-stage quantitative division framework for human driving attention and confirming that the core deficiency of algorithmic visual understanding is the lack of "semantic saliency extraction ability." By incorporating semantic attention into the human inspection stage, it can economically and efficiently fill the "semantic gap" of professional algorithms and the "grounding gap" of large models without relying on large-scale pre-training.