The vision-only End-to-End AI decision-making system overcomes the limitations of conventional autonomous driving technologies that rely on high-definition maps and RTK-based pre-recorded routes, which typically incur high deployment costs, maintenance complexity, and limited scalability for mass production. Using only onboard cameras, commercial-grade GPS, and vehicle dynamic states, the system supports autonomous driving from point A to point B on fixed urban routes. The End-to-End model adopts deep learning architectures to integrate multimodal perception inputs and performs multi-task prediction, including future trajectory forecasting, road geometry understanding, and appropriate control command generation.

1.Applicable vehicle speed ≤ 60 kph 

2.Applicable scenarios: Closed environments, fixed routes such as underground tunnels, suburban roads, intersections, etc. 

3.Drivable area detection: Forward > 60 m, Rearward > 55 m, Lateral > 9 m 

4.Lateral control acceleration < 4 m/s², Jerk < 5 m/s³ 

5.Longitudinal control speed error < 3 kph 

6.Future Trajectory Visualization >2 s

This technology can be applied to transportation shuttles in closed or semi-closed environments and fixed-route scenarios.

This technology eliminates the need for high-definition maps, significantly reducing system deployment and maintenance costs. The pure vision approach simplifies sensor types and data fusion complexity, thereby enhancing overall system reliability and maintenance convenience.