Thrixel for Physical AI
Turn real-world inputs and specifications into structured 3D assets, complete environments, and simulation data for physical AI.
Reconstruct an existing environment or generate a structured scene from a specification.
From real-world inputs
Photos and video become 3D assets and multi-object environments with triangle meshes, PBR materials, and separate semantic parts. USDA output carries the structure and physics needed for simulation workflows.
From a scene specification
Generate full structured scenes with simulation-ready assets and corresponding task definitions. Vary assets and layouts using real-world distributions, and produce robot trajectories for imitation learning.
An environment with articulated objects, physics materials, and collision geometry, ready for interaction in simulation.
Articulated parts model how doors, lids, and other components move and interact.
Physical surface properties support contact and motion within the simulation.
Collision geometry gives objects a physical presence for manipulation and environment interaction.
Assets carry information about their components, contact surfaces, and grasp locations.
Separate meshes identify meaningful components, so a handle, door, and cabinet body can each participate in a task.
Thrixel authors grasp locations for robot manipulation, with grasp-and-lift checks in simulation.
Mesh colliders and USD Physics schemas describe the contacts and constraints used during interaction.
The reconstruction workflow connects appearance, geometry, and physical behavior.
Start with imagery of a single object or an entire scene. Reference inputs and dimensions guide the reconstruction.
Create geometry, PBR textures, separate meshes for semantic parts, and simulation-ready physics in a structured asset.
Test and calibrate the asset in simulation, checking its behavior through contact, articulation, and manipulation.
Procedural variation expands an environment into multiple configurations based on real-world distributions. Structured scenes retain the assets, physics, and task definitions needed for simulation.
Robot trajectories capture interactions with the environment. Semantic parts, authored grasp points, and validated physics support the manipulation tasks used to produce imitation data.
Share the reference inputs, target environment, and robot tasks. The team can discuss reconstruction, scene generation, and simulation-data requirements.
Contact Thrixel