✦
Summary
PhysForge is a framework for generating physics-grounded 3D assets for interactive virtual worlds and embodied AI environments. It addresses a limitation of many 3D generation systems: visually plausible assets often lack the physical attributes needed for interaction, such as material behavior, functional parts, kinematic constraints, and simulation-ready structure. PhysForge focuses on assets that can be acted on, not just viewed.
Key Features
Generates 3D assets with physical properties for interaction and simulation.
Uses hierarchical physical blueprints covering material, function, and kinematics.
Employs a vision-language model as a physical architect for asset planning.
Builds on PhysDB, a large dataset with multi-tier physical annotations.
Targets interactive virtual worlds rather than only static 3D visualization.
Supports embodied AI and robotics simulation use cases.
Produces assets designed around functional logic and physical constraints.
Includes demos for virtual-world and robotic-simulation scenarios.
About PhysForge
The system uses a two-stage design supported by PhysDB, a large-scale dataset of assets annotated with hierarchical physical information. First, a vision-language model acts as a physical architect and plans a hierarchical blueprint describing material, function, and kinematics. Then a physics-grounded generation stage turns that blueprint into 3D assets suitable for virtual-world interaction, robotic simulation, and embodied reasoning.
PhysForge is useful for teams building simulation environments, game worlds, robot training scenes, or interactive digital twins where static mesh quality is not enough. By encoding how an object should behave, move, and respond to interaction, it helps close the gap between generative 3D content and physically meaningful simulation assets. The product is best viewed as a research tool for function-aware 3D asset generation.