nnsriram97/phyco_kubric

PhyCo-Sim: Synthetic Physics Simulation Dataset

This dataset accompanies the paper:

PhyCo: Learning Controllable Physical Priors for Generative Motion
Sriram Narayanan, Ziyu Jiang, Srinivasa G. Narasimhan, Manmohan Chandraker
CVPR 2026
Project Page | arXiv | PDF

PhyCo learns fine-grained, continuously controllable physical priors — friction, restitution, deformation, and applied force — from synthetic simulations, enabling physically grounded video generation without a simulator at inference. This dataset provides the training foundation for that system.

---

Overview

PhyCo-Sim contains ~90,000 photorealistic simulation videos across 9 physics scenario types, rendered using Kubric (Blender + PyBullet). Each video comes with dense ground-truth annotations: instance segmentation, depth, and full per-object physics metadata. Physical properties are systematically varied across scenes to provide rich, controlled supervision.

Scenario	Description	Varied Properties	Size
ball_drop_v2	Rigid ball falling onto a platform and bouncing	Bounciness (restitution)	1.3 GB
ball_drop_soft_v4	Deformable elastic ball falling onto a surface	Deformation stiffness	9.0 GB
ball_drop_v3	Multiple rigid balls (3–5) dropping simultaneously	Bounciness (restitution)	1.9 GB
ball_wall_collision	Ball rolling into a wall and bouncing back	Bounciness (restitution)	2.4 GB
cube_deform_soft_v2_noeff	Rigid ball impacting a soft elastic cube	Deformation stiffness	2.2 GB
friction_slide_flat_v2	Rectangular brick sliding on a flat surface	Friction, slide direction	4.2 GB
friction_slide_flat_force_v3	Brick sliding under an applied force	Force magnitude, direction	2.1 GB
jenga_force	Force applied to a single block in a Jenga tower	Push direction	3.0 GB
pool_table_force	Force applied to a ball on a billiards table	Force, direction, friction, bounciness	1.4 GB

Total compressed size: ~27.5 GB

---

Dataset Structure

Each scenario folder contains:

<scenario_name>/
├── YYYY-MM-DD.tar.gz          # Video data, batched by generation date
├── common_caption_cosmos.pt   # Pre-computed T5-XXL text embeddings (PyTorch)
├── common_caption_cosmos.txt  # Plain-text caption describing the scenario
├── props_of_interest.json     # Physical properties varied in this scenario
├── fg_bg_id.json              # (some scenarios) Foreground/background seg IDs
└── data_stats_json.tar.gz     # (some scenarios) Per-property distribution stats

Note: friction_slide_flat_v2 includes multiple caption variants for different camera viewing angles (common_caption_cosmos_down.pt, common_caption_cosmos_left.pt, etc.).

Inside each .tar.gz

Extracting a .tar.gz yields date-stamped folders of individual simulation samples:

YYYY-MM-DD/
└── <hex_id>/                  # UUID per video
    ├── rgba.mp4               # RGB video
    ├── depth.mp4              # Depth map (grayscale, encoded as video)
    ├── segmentation.mp4       # Instance segmentation masks (color-coded)
    ├── metadata.json          # Full physics and rendering metadata
    ├── animation_data.pkl     # Per-frame object trajectories (position, rotation)
    └── force_annotated_00000.jpg   # (force scenarios only) First frame with force arrow overlay

---

Video Specification

Property	Value
Resolution	768 × 432 (16:9)
Frame rate	24 fps
Duration	4 seconds (98 frames)
Video codec	H.264

---

Annotations

metadata.json schema

Each sample's metadata.json contains full simulation and rendering state. Key fields:

Per-object data (object_data):

• position — [x, y, z] world-space coordinates
• quaternion — [w, x, y, z] rotation
• scale — [x, y, z] scale factors
• mass — mass in kg
• friction — friction coefficient
• restitution — bounciness (0 = inelastic, 1 = perfectly elastic)
• segmentation_id — integer ID matching the segmentation video
• segmentation_color — [R, G, B] color in the segmentation mask (0–255)
• color — object surface color [R, G, B] in [0, 1]
• metallic, roughness, specular — PBR material properties
• Soft body only: use_neo_hookean, Neo-Hookean mu/lambda, spring stiffness/damping/bending parameters

Scene-level data:

• simulation_type — scenario identifier string
• hdri_id, hdri_rotation — environmental lighting
• ground_texture, platform_texture, platform_name — surface materials
• camera_diversity — camera position and framing metadata
• segmentation_color_map — full color → object ID mapping
• depth_of_field — focus distance and aperture parameters
• rendering_efficiency — frame settling info and optimization metrics

Force datasets additionally include the force vector (magnitude + direction) in world space, and its projection into image space.

Segmentation

segmentation.mp4 encodes instance masks: each object is assigned a unique solid color, with background as black. The color → object ID mapping is in metadata.json under segmentation_color_map.

Depth

depth.mp4 encodes linearized scene depth as a grayscale video. For decoding details (min/max depth range, scale factor), refer to the camera parameters in metadata.json.

Animation trajectories

animation_data.pkl is a pickled Python object containing per-frame arrays of object positions and rotations throughout the simulation. Useful for trajectory prediction tasks.

---

Captions and Text Embeddings

Each scenario ships with:

• common_caption_cosmos.txt — a single natural-language caption describing the physics scenario
• common_caption_cosmos.pt — the corresponding T5-XXL embedding (PyTorch torch.load)

These are scenario-level (not per-video) and were used to condition the Cosmos-Predict2 video diffusion backbone in PhyCo.

---

Generation

Videos were generated using a custom pipeline built on Kubric with:

• Physics engine: PyBullet (rigid and soft body dynamics)
• Renderer: Blender 3.x Cycles (GPU-accelerated)
• Soft bodies: Tetrahedral VTK meshes with Neo-Hookean elasticity
• Assets: KuBasic primitives, custom URDF/GLB models
• Lighting: Randomized HDRI environments with varied ground/platform textures

---

Citation

If you use this dataset, please cite:

@inproceedings{narayanan2026phyco,
  title     = {PhyCo: Learning Controllable Physical Priors for Generative Motion},
  author    = {Narayanan, Sriram and Jiang, Ziyu and Narasimhan, Srinivasa G. and Chandraker, Manmohan},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
  eprint    = {2604.28169},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

---

License

This dataset is licensed under the CC BY-ND 4.0. You are free to share and redistribute the material in any medium or format for any purpose, including commercially, as long as you give appropriate credit and do not distribute modified versions. The underlying Kubric framework is Apache 2.0 licensed. See the Kubric repository for details.