PhysForge builds 3D objects with functional physics inside

Current 3D generators can spit out gorgeous cars, chairs, and teapots in seconds. The problem begins the moment you drop them into a simulation environment: they are basically solid, lifeless blocks of plastic with zero interactive logic.

Roboticists and game developers spend endless hours slicing up meshes, rigging bones, and hand-labeling masses just to make an object functional. A new project wants to automate that entire headache away.

Researchers just introduced PhysForge, a generative framework specifically designed to build physics-grounded 3D assets that are simulation-ready right out of the box.

Blueprints first, pixels second

Instead of forcing one model to imagine geometry and functional physics simultaneously, PhysForge decouples the job into two distinct stages.

First, a vision-language model acts as a physical architect. Looking at an input, it plans a hierarchical physical blueprint that outlines parts, bounding boxes, joint kinematics, materials, and mass parameters.

In the second stage, a physics-grounded diffusion model takes that blueprint and turns it into real 3D geometry. It uses a new mechanism called KineVoxel Injection to bake kinematic constraints directly into the volumetric generation process.

To make this work, the authors built PhysDB, a massive dataset containing 150,000 assets equipped with four-tier physical annotations.

The resulting objects come with working joints, correct collision bounds, and mechanical logic. A robotic arm in a simulator can immediately grab a cabinet handle and pull it open without any manual rigging.

Why it matters

We think the smart move here is acknowledging that visual appearance and mechanical physics require fundamentally different reasoning paths.

By having a language model plan real-world common sense before diffusion kicks in, PhysForge avoids the classic trap of producing shiny objects that immediately break inside a physics engine.

For embodied AI teams, interactive data has always been the primary bottleneck. If your virtual agents need to practice opening thousands of different microwaves or toolboxes, hand-rigging each model simply does not scale. PhysForge could act as a high-throughput asset engine for sim-to-real pipelines.

The real test will be generalization. While 150,000 annotated objects is a strong start, reality is full of weird latches, deformables, and intricate multi-joint mechanisms. Still, moving generative 3D from mere eye candy to functional machinery is exactly where the industry needs to go.