Artisan: A Multi-Agent System for Physically Grounded Articulated 3D Asset Generation

From "Looks Right" to "Moves Right": A New Goal for 3D Generation

For a long time, 3D asset generation research has focused primarily on visual fidelity — precise geometry and realistic textures were considered hallmarks of quality. But for downstream applications like robot simulation and embodied AI training, simply "looking right" is nowhere near enough. Whether a cabinet drawer can slide open smoothly, whether a chair will topple under gravity, whether mass and material parameters are accurate — these are the properties that determine whether an asset can actually be used in physical simulation.

A new paper introduces Artisan, a system that shifts the generation target from "reproducing how an object looks" to "capturing how an object moves and functions." The research team calls this direction "physically grounded articulated 3D asset generation" — a significant paradigm extension over existing approaches.

Artisan System Overview

Artisan: A Multi-Agent Pipeline with Physical Simulation Feedback

Artisan builds on the prior method Articraft, with its core innovation being a multi-agent pipeline enhanced with physical simulation feedback. This design draws inspiration from research in procedural generation and agentic generation.

Unlike traditional pipelines that focus solely on geometry and appearance, Artisan's generation process covers a broader range of physical properties:

Physically plausible articulation: Ensuring movable parts operate according to real-world logic
Collision-free motion: Parts do not interpenetrate during movement
Stability under gravity: Objects do not abnormally topple after placement
Accurate mass and material properties: Providing credible physical parameters for simulation

By incorporating physical simulation results as feedback into the generation loop, the multi-agent system can continuously correct outputs that violate physical laws — making the final assets truly "usable" rather than merely "presentable."

A multi-agent pipeline is an architectural pattern that decomposes complex tasks across multiple specialized agents working collaboratively. In the context of 3D asset generation, different agents can each handle sub-tasks such as geometric modeling, joint parameter assignment, material property attribution, and physical compliance verification, passing intermediate results to one another through structured interfaces. Physical simulation feedback serves as the critical closed-loop mechanism — after each generation round, the system imports the asset into a physics engine (such as MuJoCo or Isaac Sim) to run simulation, detecting violations like collision interpenetration and gravitational instability, then propagating error signals back to the responsible agents for correction. This iterative "generate–simulate–correct" cycle essentially transforms physical laws from a post-hoc evaluation criterion into an online constraint within the generation process, allowing the model to use the simulation engine itself as a source of "physical priors" rather than relying on manually annotated physical labels.

A Unified Benchmark: Four Dimensions for Evaluating Physical Usability

Measuring the quality of such assets clearly requires more than appearance-based metrics. The research team constructed a unified benchmark that evaluates assets across four key dimensions:

The Four Evaluation Dimensions

Spatial fidelity: The accuracy of the generated asset's geometry and structure
Physical plausibility: Whether physical behaviors such as collision and stability are credible
Downstream task performance: The asset's usability in actual simulation tasks
Overall quality: A comprehensive evaluation

Benchmark Data Composition

The benchmark integrates data from multiple sources at a representative scale:

100 reference assets from ArtVIP and LightWheel
50 stability evaluation prompts
50 objects with reference mass values from ABO

Each component comes with its own evaluation protocol and metrics, covering spatial, physical, and task-level dimensions. This breadth ensures that evaluation is no longer skewed toward a single "looks good" standard, but instead aligns with real-world application requirements.

Experimental Results: Lowest Collision Rate, Highest Stability Rate

The research team conducted extensive evaluations across more than ten representative models. Results show that Artisan achieves the lowest collision rate and highest stability rate among all evaluated methods, while maintaining competitive spatial fidelity.

This combination is particularly meaningful: it demonstrates that Artisan does not trade geometric accuracy for physical performance, but instead strikes a strong balance between the two. For robot learning, virtual environment construction, and similar scenarios, a low collision rate means fewer interpenetration anomalies during simulation, while a high stability rate ensures that objects behave intuitively within virtual worlds.

Implications for the 3D Generation Field

Artisan's value extends beyond being a specific system — together with its accompanying benchmark, it charts a clear direction: advancing 3D generation from "physically plausible" to "functionally usable."

As demand grows for embodied AI and robot simulation, the "physical properties" of 3D assets are shifting from optional extras to hard requirements. Mass, material, articulation, collision, and stability — dimensions once easily overlooked — are becoming new benchmarks for generation quality. The technical approach of combining multi-agent systems with physical simulation feedback offers a replicable framework for future research, while the introduction of a unified benchmark lays the groundwork for fair comparisons across different methods.

It is foreseeable that future 3D generation evaluation will increasingly emphasize "can this actually be used" rather than stopping at visual quality. Artisan's work represents a solid step toward that goal.

Embodied AI refers to a research paradigm in which an agent must perceive, plan, and act through continuous interaction with the physical world (or its high-fidelity simulation) — fundamentally different from models trained solely on offline datasets. In embodied AI training pipelines, the physical properties of 3D assets directly affect the quality of policy learning: if a drawer's resistance coefficient is distorted or an object's center of mass is misplaced, manipulation strategies learned in simulation will struggle to transfer to real environments — the so-called "sim-to-real gap." For this reason, systems capable of generating assets with accurate physical properties represent infrastructure-level work for closing this gap, with significance far exceeding that of a content creation tool. The "functionally usable" standard that Artisan pursues is essentially about providing a reliable data foundation for large-scale simulation training in embodied AI.