Beyond Rule-Based Navigation: The Future of Self-Driving with Conditional Imitation Learning and End-to-End Driving

Abstract

Autonomous driving has traditionally evolved based on rule-based architectures; however, such approaches exhibit limited flexibility in complex urban environments and unexpected situations, and suffer from performance degradation due to error propagation across modules. To address these limitations, end-to-end (E2E) imitation-learning methods have been proposed, yet their reliance on simple regression structures restricts their ability to incorporate driving intent and contextual information. This study proposes an E2E autonomous driving model, termed ViewSelective-CIL, built upon Conditional Imitation Learning (CIL). The proposed model enhances input feature representation by integrating multi-view RGB images with semantic segmentation generated by YOLOPv2. Furthermore, Vision Transformer–based self/cross-attention mechanisms and a ViewWeightGater module dynamically compute weights for left, front, and right viewpoints according to high-level commands (HLCs) and road complexity. The model also encodes speed sequences using a GRU and applies a Temporal Transformer to incorporate spatiotemporal context, while a branch-specific fully connected architecture conditioned on HLC minimizes interference between control policies. Trained on 14,873 sequences collected from CARLA Town05, the proposed model achieved an autonomous driving success rate exceeding 95% across all routes and demonstrated attention-distribution patterns similar to human drivers. The results empirically show that the proposed approach enhances both interpretability and stability in E2E autonomous driving.