Datasets:
jais
Synthetic multi-object rigid-body videos generated with
Kubric (PyBullet physics + Blender/Cycles
rendering, GPU-rendered). Scenes contain free rigid bodies with sampled per-object
physical properties; the exact object-count and parameter ranges are recorded in
config.yaml. The subject starts with a planar velocity; the other bodies start
at rest. Physical parameters may follow visible material priors or be sampled
independently. Camera and appearance are shared across branches and vary per scene.
Each instance has three branches sharing one visual scene and one camera:
| branch | description |
|---|---|
factual |
subject launched with the sampled initial velocity |
counterfactual |
identical initial scene; subject heading changes at the logged intervention step, preserving instantaneous speed |
subject_removed |
identical initial scene; subject removed at the logged intervention step |
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.
Realized intervention metadata distinguishes frame-zero velocity from the mid-clip
command. Historical data may contain first-target failures; metadata repair does
not change their original trajectories.
Build summary
{
"branches": [
"factual",
"counterfactual",
"subject_removed"
],
"instances": 10000,
"metadata_repair": {
"counterfactual_validation": "full_branch_log",
"factual_and_removal_validation": "shared_prefix_and_original_array_preservation",
"instances_verified_this_operation": 10000,
"source_revision": "761198b5d342801a5211a2e97a23239765586fb4"
},
"object_counts": {
"4": 3205,
"5": 3341,
"6": 3454
},
"render_devices": [
"GPU"
],
"rendered": 10000,
"splits": {
"test": 1019,
"train": 7968,
"val": 1013
},
"updated_at": "2026-11-15T08:37:34.075576+00:00"
}
Reproduce
git clone https://github.com/p1neapplechoco/kubric
cd kubric
python scripts/build_velocity_dataset.py --output out --seed 2323 --count 10000 --require-gpu
python scripts/publish_velocity_dataset.py --output out --repo-id <user>/<name>
Metadata repair provenance
Metadata was reconciled for 10,000 historical instances against source revision 761198b5d342801a5211a2e97a23239765586fb4. Original scene IDs, physics trajectories, ground truth, and videos are unchanged. The counterfactual log determines the executed heading change; factual and subject-removed tracking receive shared-prefix checks and byte-exact preservation of their original arrays. Historical collision outcomes, including failures of the newer first-target rule, are retained. The historical sampling configuration remains in config.yaml.
Fast prepared-data initialization
The verified prepared cohort contains 7,700 scenes, with 6,147 training, 772 validation and 781 test scenes. The raw catalog contains 10,000 scenes. Prepared edits and embedded instance metadata are aligned with active simulation-log manifests at raw metadata revision 334152c810ee92309bdfc042c015eaaf564ec550.
The initializer loads 8,481 GT/T6 state files using four archive workers and resumes completed files. The prepared archive payload is 806,656,000 bytes. Verified prepared release.
import runpy
from huggingface_hub import hf_hub_download
revision = "b144cd9d1b61bf6945717eaf52afaee2b0324588"
path = hf_hub_download(
"pineapplechoco/jais-physical-graph-work", "setup/init_data.py",
repo_type="dataset", revision=revision,
)
initializer = runpy.run_path(path)
summary = initializer["initialize"](
"data/jais-prepared", "data/downloads", revision=revision,
)
A measured Molab run initialized this cohort in 26.43 seconds, with 0.47 seconds for a warm resume. The earlier prepared release and serial downloader took 68.75 seconds in the same session.
Generation v2 settings
New generation runs use 4–10 dynamic objects and requested factual depth of 1–min(9, object_count − 1) edges. Counterfactual first-contact targets are checked against the actual simulation contacts. Mixed coupling draws coupled and independent scenes with equal probability before QC, with material coupling applied to friction and restitution. The numeric mass and damping ranges are independently sampled.
The 10,000-scene historical cohort has observed object counts of 4–6 and an archived configuration at config.yaml. Generation v2 configuration, pinned code and parameter semantics, machine-readable profile.
- Downloads last month
- 1,276