Research map / Video / Video generation
Epona: Autoregressive Diffusion World Model for Autonomous Driving
Kaiwen Zhang, Zhenyu Tang, Xiaotao Hu, Xingang Pan, Xiaoyang Guo, Yuan Liu, Jingwei Huang, Li Yuan, Qian Zhang, Xiao-Xiao Long, Xun Cao, Wei Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint video prediction with shared latent to trajectory planning (ablation) | Epona · NAVSIM test | PDMS: 78.1 -> 86.2 | Improved |
| add temporal-aware DCAE decoder (ablation) | Epona · NuPlan test | FVD40: 100.11 -> 74.88 | Improved |
| add temporal-aware DCAE decoder (ablation) | Epona · NuPlan test | FVD10: 52.95 -> 50.77 | Improved |
| add chain-of-forward training (ablation) | Epona · NuPlan test | FID over long generation | Improved |
| increase condition frames from 2 to 10 (ablation) | Epona · NuPlan test | FVD40: 103.70 -> 74.88 | Improved |
| increase condition frames from 5 to 10 (ablation) | Epona · NuPlan test | FVD25: 71.28 -> 61.46 | Improved |
| increase condition frames beyond 10 (ablation) | Epona · NuPlan test | compute | Hurt |
| autoregressive diffusion world model versus fixed-length video diffusion | Epona vs Vista · NuScenes val | FVD: 89.4 -> 82.8 | Improved |
| autoregressive diffusion world model versus fixed-length video diffusion | Epona vs Vista · NuScenes val | FID: 6.9 -> 7.5 | Hurt |
| planning from the world model without perception inputs versus end-to-end planners | Epona vs DRAMA · NAVSIM test | PDMS: 85.5 -> 86.2 | Improved |
| continuous diffusion trajectories versus discrete tokens | Epona vs DrivingGPT · NAVSIM test | PDMS: 82.4 -> 86.2 | Improved |
| no extra supervision versus annotated supervision for planning | Epona vs UniAD · NuScenes planning | average L2 (m): 1.03 -> 1.25 | Hurt |
Later papers that compared with this method
2 later accepted papers report Epona as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- DriveLaW: Unifying Planning and Video Generation in a Latent Driving World (CVPR 2026 · better on 7 of 8 shared results)
- DriveVA: Video Action Models are Zero-Shot Drivers (ECCV 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- E-Motion: Future Motion Simulation via Event Sequence Diffusion (NeurIPS 2024 · similarity 0.88)
- Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving (CVPR 2024 · similarity 0.87)
- DiffusionDrive: Truncated Diffusion Model for End-to-End Autonomous Driving (CVPR 2025 · similarity 0.87)
- SAMPO: Scale-wise Autoregression with Motion Prompt for Generative World Models (NeurIPS 2025 · similarity 0.86)
- MaskGWM: A Generalizable Driving World Model with Video Mask Reconstruction (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
