SIMART: Turning Static 3D into Sim-Ready Robots

Static 3D generation is obsolete; physical agents now need articulated objects that can actually move and interact with physics engines.

New research introduces SIMART, a unified AI framework that instantly decomposes monolithic 3D models into complex, sim-ready articulated assets without the error-prone multi-step pipelines of the past. The system uses a Sparse 3D VQ-VAE to analyze geometry and kinematics simultaneously, slashing token counts by 70% compared to dense voxel methods. This efficiency allows high-fidelity assembly of multi-part objects directly from raw meshes, enabling robots to simulate grasping, lifting, and manipulating complex machinery in real time.

The result is a single-stage process where the model understands object parts and their physical constraints at once, bridging the gap between static generation and dynamic robotic simulation. Developers can now generate ready-to-test assets that work immediately in physics environments, removing the tedious manual tagging of joints and hinges required by previous tools.

Source: SIMART: Decomposing Monolithic Meshes into Sim-ready Articulated Assets via MLLM by Zhang et al., arXiv:2603.23386