A Survey of End-to-End Driving: Architectures and Training Methods

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving technology is a rapidly growing field with significant academic and industrial interest.
• Machine learning has been extensively applied to autonomous driving, primarily focusing on perception tasks.
• End-to-end approaches, utilizing a single neural network for the entire driving pipeline, represent a newer direction.

Purpose of the Study:

• To provide a comprehensive review of end-to-end deep learning approaches in autonomous driving.
• To analyze learning methods, input/output modalities, network architectures, and evaluation strategies in this domain.
• To discuss critical challenges, specifically interpretability and safety, associated with end-to-end driving systems.

Main Methods:

• Systematic literature review of end-to-end autonomous driving research.
• Categorization and analysis of various learning techniques, data modalities, and network designs.
• Examination of common evaluation metrics and benchmarks used in the field.

Main Results:

• Identified diverse learning methods, input/output modalities (e.g., camera images, LiDAR), and neural network architectures.
• Highlighted common evaluation schemes and metrics used to assess system performance.
• Confirmed that interpretability and safety remain significant challenges for current end-to-end driving systems.

Conclusions:

• End-to-end deep learning offers a promising, albeit challenging, paradigm for autonomous driving.
• Future research should focus on addressing safety and interpretability concerns.
• A proposed architecture combining successful elements from existing systems is presented as a potential advancement.