Tsinghua Team Reveals Key Differences in Visual Attention Between Humans and Autonomous Driving Algorithms

In a study published in *npj Artificial Intelligence* in February 2026, the team from the Institute for AI Industry Research (AIR) at Tsinghua University conducted the first dual-track experiment combining "human eye-tracking with algorithmic comparison" to systematically analyze the fundamental differences in visual attention between human drivers and autonomous driving algorithms. The research introduced a three-stage quantitative framework for human driving attention and revealed that the core limitation of current intelligent driving algorithms lies in their lack of "semantic saliency extraction capability." The team further demonstrated that integrating the semantic attention mechanism observed in humans during the inspection phase into algorithms can effectively bridge the "semantic gap" of specialized models and the "grounding gap" of large models—without relying on large-scale pretraining—offering a novel pathway to enhance the safety of autonomous driving systems.