Procedural Generation of Articulated Simulation-Ready Assets

TL;DR

Robotics simulation relies on diverse articulated objects with accurate joints, which existing asset datasets struggle to provide. The paper presents Infinigen-Articulated, a Blender-based procedural toolkit that generates 18 categories of articulated assets with ground-truth kinematics, photorealistic materials, and automated export to URDF, USD, and MJCF for simulation. It enables fine-grained control over joint parameters, scalable detail, and consistent metadata, facilitating rigorous perception and policy learning. Across movable-part segmentation, reinforcement learning generalization, and sim-to-real transfer tasks, assets generated by Infinigen-Articulated improve data diversity and policy performance, bridging the sim-to-real gap.

Abstract

We introduce Infinigen-Articulated, a toolkit for generating realistic, procedurally generated articulated assets for robotics simulation. We include procedural generators for 18 common articulated object categories along with high-level utilities for use creating custom articulated assets in Blender. We also provide an export pipeline to integrate the resulting assets along with their physical properties into common robotics simulators. Experiments demonstrate that assets sampled from these generators are effective for movable object segmentation, training generalizable reinforcement learning policies, and sim-to-real transfer of imitation learning policies.

Figures (6)

• Figure 1: Procedurally generated assets for 18 common articulated object categories.
• Figure 2: We first create procedural assets using Blender's Geometry Nodes feature. These assets are then transpiled, and we have the ability to define custom distributions for the procedural parameters (including materials) along with its dynamics properties. We sample parameters to generate the asset and pass both the asset and its kinematic tree to the exporter, which recursively builds a simulation-ready articulated object.
• Figure 3: Comparison of asset parameter distributions. The left column shows the default distribution used for handle length and handle-to-hinge distance. The right column shows an expanded distribution where these parameters are varied over a wider range.
• Figure 4: Examples of articulations for small, but essential parts (rows 1 and 2), occluded objects (rows 3 and 4), and double-jointed objects (row 5).
• Figure 5: Left: Environment setups for each task. Right: Average success-once performance over 5 seeds with bounds showing $\pm0.5$ standard deviation.