Autonomous Driving Held Back by System Integration, Not AI Algorithms

In recent years, the accuracy of perception algorithms in the autonomous driving industry has improved by nearly 100 times, with large models now capable of identifying objects—such as pedestrians, plastic bags, or traffic cones—in just a few milliseconds. Meanwhile, computing costs have dropped significantly, and chips costing only a few dozen dollars can now run full visual perception models. Although the AI field continues to receive substantial investment in capital, talent, and technology—and has achieved remarkable progress—the industry as a whole still faces bottlenecks. The issue does not lie with AI itself, but rather with other components of the vehicle system that fail to effectively integrate and collaborate with these advanced algorithms.