Articraft turns 3D asset generation into a coding problem

Most generative 3D tools on the market today share the same frustrating limitation: they spit out a single, static chunk of mesh. You can prompt a visually stunning car or cabinet in seconds, but you cannot open the trunk, pull the drawer, or flex a hinge. If you want to drop these assets into a robotics simulator or a game engine, someone still has to manually rig the joints, separate the parts, and define the mechanical limits.

Researchers open-sourced Articraft, an agentic system and accompanying dataset that tackles this bottleneck by rethinking the entire generation pipeline.

Turning 3D assets into executable code

The clever twist behind Articraft is that it doesn't ask an AI to directly hallucinate complex 3D meshes or fiddle with raw point clouds. Instead, it turns asset creation into a programming challenge.

The system sets up an agentic harness where an LLM writes code against a domain-specific SDK. It supports models like GPT-5.5, Claude, or Gemini without being locked to a single provider. The harness gives the model a restricted workspace, freeing it from the headache of hand-authoring URDF files or wrangling messy environments.

Inside this environment, the agent writes a program that specifies individual parts, composes their geometry, and defines how joints, sliders, and hinges connect them. Crucially, the model also writes tests to sanity-check the resulting object. The harness runs the program, validates the physics and connectivity, and returns structured feedback so the agent can debug its own mistakes.

So far, the results are remarkably versatile. From simple text prompts, Articraft has generated multi-joint desk lamps, three-bar linkage chains, functional mechanical keyboards with separate keycaps, and even multi-stage jet engines.

Alongside the system, the team released Articraft-10K, a curated dataset of over 10,000 articulated 3D assets spanning 245 categories, ranging from everyday tools and furniture to complex machinery.

Why it matters

We think Articraft's approach is genuinely smart because it sidesteps the fundamental weakness of standard 3D diffusion models: multi-part mechanical reasoning. Diffusion models are great at shapes and textures, but they consistently struggle with rigid constraints, joint axes, and moving components. By leaning into code generation, Articraft converts a tricky spatial synthesis problem into a structured logic problem that modern LLMs are already great at solving.

For robotics and simulation teams, this is a massive operational win. Embodied AI needs millions of interactive, articulated objects to teach robot arms how to open doors, turn knobs, and manipulate real-world clutter. Hand-rigging those environments has kept simulation data scarce and expensive. Articraft cracks that bottleneck wide open, offering a realistic path toward on-demand synthetic training worlds.

That said, the code-first approach has its boundaries. It excels at geometric, man-made objects with clear mechanical logic, but it won't easily describe organic, freeform shapes. The real breakthrough down the road will likely happen when code-driven joint assembly merges with high-fidelity visual surface generators.