Reproduce “Auto-JEPA: A Latent World Model of Continuous Intent 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
Jiwei Yang, Zhengxian Chen, Chaosheng Huang, Jun Li
A latent world model named Auto-JEPA from Tsinghua University predicts continuous future ego-motion intent for end-to-end autonomous driving. The system facilitates trajectory retrieval and selection by selectively attending to planning-relevant dynamic agents, achieving a 91.3 PDMS on NAVSIM v1 and 89.1 EPDMS on NAVSIM v2.
Key insights
- High-quality planning can be achieved without dense future-world modeling by focusing on predicting a compact, continuous future ego-trajectory latent.
- The predicted future ego-trajectory latent can directly serve as a planner's retrieval key, demonstrating a novel operational role for predicted latents in real-time inference.
- The system implicitly learns to selectively attend to dynamic agents critical for planning decisions without explicit object-level annotations or interaction labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.