Epona: Autoregressive Diffusion World Model for Autonomous Driving
TLDR
Epona is an autoregressive diffusion world model for autonomous driving enabling long-horizon video prediction and trajectory planning with state-of-the-art performance.
Reasoning
The paper introduces a novel autoregressive diffusion approach with decoupled spatiotemporal factorization and chain-of-forward training, achieving strong results on benchmarks. However, the method is domain-specific to autonomous driving and may still face error accumulation in very long predictions.
Read-first score
Read-first score 74.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.
Field roles
Rank sensitivity
Stability: volatile; rank range: 71.
Keyword Scores
Deep Analysis
Innovations
- Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation
- Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework
- Chain-of-forward training strategy to address error accumulation in autoregressive loops
Methodology
Epona is an autoregressive diffusion world model that models localized spatiotemporal distributions by decoupling temporal dynamics from future world generation and integrating modular trajectory and video prediction. It employs a chain-of-forward training strategy to mitigate error accumulation in autoregressive loops, enabling high-resolution, long-duration video generation.
Key Results
Epona achieves state-of-the-art performance with a 7.4% improvement in FVD and minutes longer prediction duration compared to prior works, and the learned world model serves as a real-time motion planner outperforming strong end-to-end planners on NAVSIM benchmarks.