Reproduce “ActiveAD: Planning-Oriented Active Learning for End-to-End Autonomous Driving” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Han Lu, Xiaosong Jia, Yichen Xie, Wenlong Liao, Xiaokang Yang, Junchi Yan
ActiveAD introduces a planning-oriented active learning framework for end-to-end autonomous driving, incorporating criteria tailored to driving performance and safety. The method enables an end-to-end autonomous driving model to achieve comparable or even superior planning performance using only 30% of the nuScenes dataset compared to training on the entire labeled dataset.
Key insights
- Strategic data selection can be more valuable than simply collecting and annotating vast quantities of undifferentiated data for end-to-end autonomous driving (E2E-AD) planning.
- Designing active learning criteria that directly align with the planning objective (Displacement Error, Soft Collision, Agent Uncertainty) significantly improves label efficiency and planning performance in E2E-AD.
- Prioritizing initial data diversity through meta-information like 'Ego-Diversity' (weather, driving commands, speed) provides a more robust foundation for subsequent active learning iterations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.