jais

Synthetic multi-object rigid-body videos generated with Kubric (PyBullet physics + Blender/Cycles rendering, GPU-rendered). Every scene has 4-6 fixed-mass bodies on a table; exactly one body (the subject) receives an initial velocity along a single heading and nothing else is actuated. Some bodies sit in the subject's corridor and get struck, others are bystanders that are never touched. Spheres roll, boxes slide, and a sphere launched without spin slides first and then rolls. Materials (and the coupled friction / restitution) are sampled per object; the camera is fixed within a clip and re-sampled between clips.

Each instance has three branches sharing one visual scene and one camera:

branch	description
factual	subject launched with the sampled initial velocity
counterfactual	same scene, the factual velocity is discarded and a new initial velocity is applied at step 0
subject_removed	same scene without the subject

Contents

manifest.jsonl                 one row per instance (split, roles, velocities, QC, ground truth summary)
dataset_summary.json           counts and render devices
config.yaml                    sampling ranges used for this build
instances/<index>/
  instance.json                full scene spec (objects, physics, materials, camera)
  ground_truth.json            contact-graph delta, hard/soft affected objects, propagation paths
  qc.json                      QC metrics: rolling / sliding labels, struck & untouched objects
  <branch>/video.mp4           RGB clip (48 frames @ 24 fps)
  <branch>/mask.mp4            colourised instance segmentation preview
  <branch>/depth.mp4           colourised depth map video preview
  <branch>/flow.mp4            colourised optical flow video preview
  <branch>/segmentation.npz    uint8 [T,H,W] instance ids (0 = floor/background)
  <branch>/depth.npz           float16 [T,H,W] depth
  <branch>/forward_flow.npz    float16 [T,H,W,2] optical flow
  <branch>/tracking.npz        per-frame poses, velocities, 2D projections, boxes, visibility, presence
  <branch>/graph.json          temporal contact graph (object-object episodes; floor episodes listed separately)
  <branch>/sim_log/            immutable per-physics-step states + contacts (manifest-hashed)

segmentation_ids / object_ids in tracking.npz are shared across the three branches, so masks and tracks line up between factual, counterfactual and subject-removed clips.

Build summary

{
  "branches": [
    "factual",
    "counterfactual",
    "subject_removed"
  ],
  "instances": 7700,
  "render_devices": [
    "GPU"
  ],
  "rendered": 7700,
  "splits": {
    "test": 781,
    "train": 6147,
    "val": 772
  },
  "updated_at": "2026-09-21T21:40:21Z"
}

Reproduce

git clone https://github.com/p1neapplechoco/kubric
cd kubric
python scripts/build_velocity_dataset.py --output out --seed 31113133311131311313131 --count 7700 --require-gpu
python scripts/publish_velocity_dataset.py --output out --repo-id <user>/<name>

Counterfactual video generation splits (derived)

A second, derived train/val/test partition used for counterfactual video generation. It is independent of the official splits in the build summary above. Scenes with index < 2000 and >= 2000 come from different generation configs, so the two index ranges are split separately (subset A and subset B) rather than pooled.

subset	scene ids	scenes	train	val	test
A	000000-001999	2000	1600	200	200
B (full)	002000-009899	5700	4560	570	570
B (balanced)	002001-009899	2000	1600	200	200

• Ratios 80/10/10, random with a fixed seed (0), stratified by the number of dynamic objects (4, 5 or 6), done independently inside each subset. Scene ids are not contiguous (7700 scenes exist, up to id 009899).
• B (balanced) is a subsample of B (full) so that subset B has as many scenes as subset A in every partition. It is drawn inside B's own train/val/test lists (stratified by number of dynamic objects, seed 100), so no scene changes partition.
• Files used per scene: instance.json, and for the factual and counterfactual branches tracking.npz, segmentation.npz and video.mp4.