The Pulse
Artisan Adds Physics Checks to AI-Generated 3D Assets
Artisan combines AI agents with feedback from a physics simulator to test whether generated 3D objects move and behave as intended. Its reported results cover collision rates, stability, human preferences and simulated robot tasks.
AI.info Team · Written with AI ·
Artisan’s best-performing configuration recorded collisions in 1.7% of single-joint motion tests and kept 98% of tested objects stable at rest, according to a paper submitted to arXiv on October 9. The system targets a persistent weakness in AI-generated 3D models: an object can look convincing while its moving parts intersect, its body tips under gravity, or its physical properties make it unsuitable for simulation.
Rather than treating appearance as the finish line, the team behind Artisan tests whether articulated objects move and behave as intended. The paper describes a pipeline that combines separate AI agents with feedback from the MuJoCo physics simulator, then evaluates the resulting assets on geometry, physical behavior, simulated robot tasks and human preference.
Three agents build, then challenge, each asset
Artisan accepts a text prompt and, optionally, reference images. A Design Agent lays out the object’s parts, joints, motion and function; a Coding Agent turns that plan into a 3D asset and program; and a separate Review Agent checks the result against the design. Keeping review separate from construction is meant to reduce the risk that one agent simply approves its own work.
MuJoCo supplies another check. The system simulates moving parts to identify collisions, recording which components intersect and how deeply, and tests whether objects stay upright under gravity. That report goes back into the coding-and-review loop, where agents can revise the geometry or physical parameters. The pipeline also assigns properties such as part mass and material, and describes interactions an object should support.
The paper’s authors, Xudong Cai, Ruijie Zhu, Haoliang Shang, Ruining Li, Marc Pollefeys, Shangzhe Wu, Andrea Vedaldi and Chuanxia Zheng, write:
“Through extensive evaluation of more than 10 representative models, Artisan achieves competitive spatial fidelity while attaining the lowest collision rates and highest stability rates among the evaluated methods.”
The physics results depend on the test
In the paper’s ArtVIP benchmark, the GPT-6 Astra version of Artisan reports a 1.7% collision rate for single-joint motion, 3.7% for double-joint motion and 13.0% for multi-joint motion. The team also reports stability rates of 98% for objects at rest and 96% in articulated states. Those results come from particular datasets and simulated checks—not a general guarantee that every generated model will behave correctly.
The project page’s GPT-5.5 comparison shows a 3.5% single-joint collision rate with physics validation and 37.5% without it; multi-joint collisions rise from 12.3% to 77.6% without physics. This comparison concerns those GPT-5.5 variants and does not establish that simulation feedback alone accounts for every difference between competing methods.
The benchmark combines 100 reference assets from ArtVIP and LightWheel, 50 prompts for stability testing and 50 ABO objects with reference masses. It also measures spatial fidelity, material and mass estimates, and how assets fare in downstream simulated tasks. The authors report a first-choice preference of 43.3% for Artisan, compared with 21.1% for Codex, in a user evaluation following object manipulation in MuJoCo.
Robot-training gains remain simulated
In downstream tests, a policy trained on a mix of native and Artisan-generated assets reached an average 84% success rate on RoboTwin tasks and 43% on RoboCasa tasks, evaluated on held-out native objects. The native-only training baseline scored 80% and 39%, respectively. Artisan-generated assets alone scored 63% on RoboTwin and 27% on RoboCasa, so the results do not show that synthetic assets replace existing training data.
Cost and time also matter for practical use. The paper reports an average API cost of $5.79 and generation time of 34.3 minutes per asset for Artisan using GPT-6 Astra. Its authors state that the assets were intended for simulation and evaluation; real-world robotic use requires additional validation. The reported gains therefore establish a stronger case for testing generated objects inside simulators—not proof that a virtual cabinet door or appliance is ready to handle a physical robot.