Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark

Towards Physically Grounded Articulated 3D Assets. A Multi-Agent Pipeline and Benchmark. By Xudong Cai and seven others, from Nanyang Technological University, ETH Zurich, and three other institutions. Published on the ninth of October. Its code repository is worked out: about three hundred and fifty source files, with a readme and a licence.

Generating articulated 3D assets has mostly meant reproducing how objects look. The aim set out here is physically grounded generation: capturing how objects move and how they function. An asset can look right and still collide with itself when a part moves, fall over under gravity, or carry the wrong mass. For an asset meant to be opened, pushed or picked up, those failures matter as much as its shape. Matching appearance is no longer enough to call the job done.

Drawing on work in procedural and agentic generation, the authors present Artisan. It extends the recent method Articraft by introducing a multi-agent pipeline that receives feedback from physical simulation. Rather than judging an asset only by how it looks, the pipeline can see how it behaves when simulated.

That feedback lets the system reach attributes beyond geometry, appearance and plausible articulation. These include motion free of collisions, stability under gravity, and accurate mass and material properties. Alongside the system, the authors build a unified benchmark covering four aspects: spatial fidelity, physical plausibility, downstream task performance, and overall quality. It contains one hundred reference assets drawn from ArtVIP and LightWheel, fifty stability prompts, and fifty ABO objects with reference masses, with protocols for evaluating a range of metrics. Each part of the benchmark lines up with a physical property: the stability prompts with standing under gravity, the objects with reference masses with getting weight right.

More than ten representative models are evaluated. Artisan achieves competitive spatial fidelity, and among the methods evaluated it has the lowest collision rates and the highest stability rates. In other words, it keeps pace on how assets look while leading on the physical properties the simulation feedback was introduced to address.

The shift being argued for is in what counts as a finished asset. Shape and appearance remain part of it, but an object that is meant to be used should also move without collisions, stand up under gravity, and weigh what it should, and the benchmark is built to check exactly those.