haodongli/LDR

Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

Please visit our Github Repo for how to use these datasets.

Structure

ID: in-distribution, OOD: out-of-distribution.

LDR/
├── train/                  # training clips (ID only)
│   ├── uniform.hdf5        # 30K clips
│   ├── parabola.hdf5       # 30K
│   ├── collision.hdf5      # 30K
│   ├── looming.hdf5        # 30K
│   ├── bouncing.hdf5       # 30K
│   └── joint_5task.hdf5    # 150K, the five single-task training sets concatenated
├── eval/                   # evaluation clips (ID + OOD)
│   ├── uniform.hdf5
│   ├── parabola.hdf5
│   ├── collision.hdf5
│   ├── looming.hdf5
│   ├── bouncing.hdf5
│   └── joint_5task.hdf5    # the five single-task eval sets concatenated
└── splits/                 # ID/OOD index lists
    ├── uniform.json
    ├── parabola.json
    ├── collision.json
    ├── looming.json
    ├── bouncing.json
    └── joint_5task.json    # the five single-task splits concatenated

BibTeX

@article{li2026learning,
  title={Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning},
  author={Li, Haodong and Liu, Shaoteng and Wang, Tianyu and Ge, Chongjian and Ji, Sihui and Zhang, Jiahan and Lin, Xin and Lu, Haolin and Lin, Zhe and Chandraker, Manmohan},
  journal={arXiv preprint arXiv:2608.09926},
  year={2026}
}