Empowering Autonomous Vehicles to Make Challenging Options in Unexpected Circumstances with Hybrid Learning

Autonomous vehicles (AVs) must navigate challenging and unexpected circumstances while guaranteeing security and competence. Prescribed rule-based classifications strive to handle the large unpredictability of virtual driving situations. In the proposed work, a novel hybrid architecture enables autonomous vehicles to make human-like choices in unexpected scenarios by using a combination of deep learning and data-driven planning techniques. The framework combines VOLOv7-based perception, multimodal transformers for fusing LiDAR, radar, and camera data, and a dual-policy approach using DAgger and Decision Transformer to obtain both sensitive and deliberate decision-making behaviors. An ensemble voting mechanism combines policy outputs to improve reliability. The proposed work is trained and evaluated using the Waymo Open Dataset and CARLA simulator. The proposed work attains a collision rate of 3.4%, route completion of 97.2%, and an average intervention frequency of 0.4.