JiecaoBoy/CASTER-scenes

This dataset accompanies the paper CASTER: Formalizing Zero-Shot Manipulation from Synthetic Videos as Constrained Optimization.

For more details, see the project page and the code repository.

Tasks are numbered in the order of Table I in the released paper.

Each scene contains its initial RGB observation (assets/rgb.png), point cloud (assets/pointcloud.ply), camera-to-robot calibration (assets/C2R.npy), and 20 generated MP4 videos under the task folder. Video filenames preserve generation seeds 0–19. All generated samples are included, regardless of task-validation outcome.

scene_1/
  assets/
    rgb.png
    pointcloud.ply
    C2R.npy
  move_banana_in_pan/
    output_seed_0.mp4
    ...
    output_seed_19.mp4
...
scene_12/