SSIL: Self-Supervised Imitation Learning for End-to-End Driving

The paper introduces Self-Supervised Imitation Learning (SSIL), a novel framework for end-to-end (E2E) autonomous driving that eliminates the need for extensive human driving data or pre-trained models. SSIL generates pseudo steering angle data using vehicle poses from LIDAR sensors, and incorporates a cross-attention-based conditioning approach (CACA) to enhance visual information processing. Experimental results on three benchmark datasets show that SSIL achieves comparable E2E driving accuracy to supervised methods, highlighting its potential to reduce reliance on large labeled datasets in autonomous driving applications.