Reproduce “GenAD: Generalized Predictive Model for 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
Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu and 6 more
This work presents GenAD, a generalized video predictive model for autonomous driving, trained on OpenDV-2K, the largest public multimodal driving dataset at over 2000 hours. The system demonstrates strong zero-shot generalization for video prediction across diverse environments and achieves efficient adaptation for action-conditioned simulation and highly competitive planning, showcasing a unified approach to learning driving scene dynamics.
Key insights
- Large-scale pre-training on diverse, real-world driving videos enables a single model to generalize across varied geographic locations, weather, and sensor configurations, moving beyond specialized models.
- A two-stage training paradigm, initializing with a general image diffusion model and then adding causal temporal modules, effectively adapts high-quality image generation to dynamic video prediction in driving scenes.
- Learning a generative 'world model' through video prediction provides rich spatiotemporal representations that are highly transferable and accelerate downstream autonomous driving tasks like simulation and planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.