Efficient and Predictive Self-Driving with Large AI Models

AI Technology: Deep Learning; Vision (image recognition, image generation, text recognition OCR, etc.).

End-to-end autonomous driving has shown promise with large AI models, yet their real-time deployment remains a challenge due to high computational costs. 

Inspired by dual-system cognitive theories, the project proposes a hybrid framework where a lightweight, fast-acting model operates in real-time while a larger, high-capacity model predicts future representations and refines decision-making. 

Unlike existing methods that process the current timestep in isolation, our approach anticipates potential failures, ensuring timely corrections without prohibitive inference delays. 

The projec team plans to validate their method in the CARLA v2 closed-loop setting using the Bench2Drive benchmark, focusing on complex driving scenarios. 

By leveraging large models for predictive intervention rather than direct control, the project enhances safety, robustness, and efficiency while reducing onboard computational demands. 

This approach not only advances autonomous driving but also provides a practical framework for integrating large models into real-world decision-making systems.