World perception models for embodied AI

From a photograph to a world you can edit, interact with and manipulate.

Understanding object structure and physical properties

Reconstruct editable, interactive 3D environments from photographs, with spatial relationships, internal structures and moving parts that support manipulation in simulation.

Choose a scene to explore its structure and motion.

Input photograph
Input photograph of the kitchen

Kitchen

Cabinet doors swing open, drawers slide out, and the microwave and oven doors reveal their interiors.

Reconstruction and part motion
Watch the scenes side by side

From object structure to physical interaction

Doors rotate about hinges. Drawers follow a sliding direction. Objects fall, collide and come to rest. Structure and physical properties shape how a scene responds to interaction.

Apples fall into a bowl; the overlay reveals the objects’ collision geometry.

Objects and parts

Color-coded views show individual objects and parts within each reconstructed scene.

Joints and motion

Hinges and sliding joints define how a part moves and its range of motion.

Physical properties

Mass, friction and collision geometry provide the simulation with the properties needed for contact and motion.

Edit objects within the scene

Change a wooden tabletop to marble or a pendant lamp to brass with a text instruction. Objects can be edited individually within the reconstructed scene.

Build a world model for your task

Tell us about your target environment and the interactions you want to simulate.

Discuss your scene

Technical foundations

From observations to actionable environments

Object reconstruction and part generation provide a structural foundation for actionable 3D environments. Scale calibration, physical properties and simulation checks prepare scenes for specific interaction tasks.

Reconstruct objects and where they belong

CUPID jointly estimates object geometry and camera pose, using image-aligned features to preserve observed detail. SPOON uses multi-view geometry to coordinate object scale, position and orientation in a coherent scene. Together, these research directions address fidelity to observations and the spatial arrangement of objects.

CUPID reconstructs a toy and its camera pose from a photograph, then renders the reconstructed toy from that pose.
CUPID: input image, reconstructed object and camera, and a view rendered from the recovered camera pose. CUPID · Fig. 1 ↗

Re-rendering from the recovered camera pose checks fidelity to visible evidence. Regions outside the photograph are completed by the model and can be checked against additional views.

Explore the scene-layout method
The SPOON pipeline: multi-view scene geometry guides object generation, selects reliable views and jointly refines objects and cameras.
SPOON uses shared geometry, view selection and joint refinement to align objects and cameras in a coherent scene. SPOON · Fig. 2 ↗

Preserve part detail and whole-object structure

STRUCTPARTS alternates between part-local and object-global coordinates to preserve fine geometry and the arrangement of parts. Changing the segmentation groups allows parts to be split or merged. Explicit parts provide a basis for assigning joints and motion limits.

STRUCTPARTS examples show how beetle and bear models can be split into smaller parts or have parts merged.
Changing segmentation groups produces the same object at different levels of part granularity. The examples show split and merged part structures. STRUCTPARTS · Fig. 7 ↗

Test structure and motion in simulation

A scene intended for robot manipulation needs to be tested in simulation. The team’s workflow brings scale calibration, part modeling, physical properties and simulation checks together, using the results to validate structure and motion.

  1. Calibrate camera and scale

    Use structural lines and objects with known dimensions as references, then cross-check the scale against other standard objects.

  2. Align objects and parts

    Align objects to shared planes such as walls, counters and cabinet fronts, while keeping doors and drawers as separate parts.

  3. Assign motion and physical properties

    Assign hinges, sliding joints and motion limits to moving parts, and estimate object mass from material.

  4. Check rendering and simulation

    Compare rendered views with the input photograph. Check stability, full part motion and interpenetration in MuJoCo, followed by an independent validation pass.