End-to-End Autonomous Driving Without Costly Modularization and 3D Manual Annotation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Current end-to-end autonomous driving (E2EAD) models often rely on supervised perception and prediction, requiring extensive 3D annotations.
• This reliance on annotations hinders data scaling and incurs substantial computational costs during training and inference.
• Existing E2EAD models can mimic modular architectures, leading to inefficiencies.

Purpose of the Study:

• To propose UAD, an end-to-end framework utilizing an unsupervised pretext task for vision-based autonomous driving.
• To overcome the limitations of supervised learning in E2EAD, specifically annotation requirements and computational overhead.
• To enhance both open-loop evaluation performance and closed-loop driving quality.

Main Methods:

• Developed UAD, an end-to-end framework featuring an unsupervised pretext task for autonomous driving.
• Introduced an Angular Perception Pretext to predict angular-wise spatial objectness and temporal dynamics without manual annotation.
• Implemented a self-supervised training strategy focusing on trajectory consistency under augmented views to improve steering robustness.

Main Results:

• UAD achieved state-of-the-art open-loop performance on nuScenes, with a 38.7% relative improvement in average collision rate over UniAD.
• Demonstrated robust closed-loop driving in CARLA, achieving a 98.5% route completion score on the Town05 Long benchmark.
• Consumed 44.3% less training resources and achieved 3.4x faster inference compared to UniAD.

Conclusions:

• UAD offers significant performance advantages over supervised E2EAD methods.
• The unsupervised approach drastically reduces data annotation needs and computational demands.
• UAD presents a more efficient and scalable solution for autonomous driving systems.