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.
Introduction
Training robots to perform complex manipulation tasks, such as opening a dishwasher or operating a toaster, requires a vast and diverse set of 3D models. Traditionally, this has been a significant bottleneck in "sim-to-real" transfer. Existing datasets like PartNet-Mobility rely on manual human annotation to define joint locations and ranges of motion. This process is not only slow and error-prone but often misses small details like buttons or internal components. Furthermore, manual datasets often lack kinematic diversity; if a robot only trains on doors with handles in a single location, it fails to generalize to real-world variations.
Infinigen-Articulated addresses these challenges by automating the creation of high-quality, simulation-ready assets with moving parts. By leveraging procedural generation, it provides perfect joint annotations and unlimited geometric variety, significantly outperforming traditional datasets in training robots for real-world tasks.
Methods
The toolkit is built on the Infinigen ecosystem, utilizing Blender’s Geometry Nodes to define objects through mathematical rules rather than manual sculpting.
- Procedural Joint Definition: Specialized node groups for revolute (hinge) and prismatic (sliding) joints are integrated directly into the procedural "recipe." This ensures that ground truth data, such as rotation axes and motion ranges, are mathematically exact by construction.
- Transpilation and Sampling: Blender node graphs are transpiled into Python code, allowing for programmatic sampling of infinite variations. Parameters such as bevel radii, dimensions, and component counts can be varied continuously.
- Simulation Integration: A dedicated pipeline exports these procedural objects into standard formats like URDF, USD, and MJCF. This includes the automatic calculation of physical properties like friction and density, and the construction of kinematic trees for simulator-ready hierarchies.
Datasets
The authors provide procedural generators for 18 common articulated object categories, ranging from kitchen appliances like ovens and microwaves to household fixtures like faucets and lamps.
In evaluations, these assets were compared against the industry-standard PartNet-Mobility (Sapien) dataset and a vision-specific subset termed Infinigen-Sim.
Key Insights and Results
A primary advantage of procedural generation is the ability to achieve dense coverage of the kinematic task space. By continuously varying parameters like handle length and its distance from the hinge, the system prevents Reinforcement Learning (RL) policies from overfitting to specific geometries.
Experimental Performance
- Computer Vision: Models trained with Infinigen assets showed superior performance in "Movable Part Segmentation" (mAP), particularly for small components like buttons that are often omitted in manual datasets.
- Reinforcement Learning: Policies trained on Infinigen-Articulated assets achieved a 2.86x average improvement in success rates over those trained on PartNet-Mobility when tested on novel objects.
- Sim-to-Real Transfer: In zero-shot real-world tests with a Franka Panda robot, the proposed method significantly outperformed the baseline. For door opening, the success rate was 18/30 compared to 3/30 for PartNet. For toaster and fridge tasks, the baseline failed completely (0/30) while Infinigen-trained policies achieved over 10/30 successes.
Conclusion and Future Work
Infinigen-Articulated demonstrates that procedural generation is a powerful tool for scaling up simulation data for robotics. While currently limited to independent joints and high-detail meshes that can impact simulation speed, future work aims to support coupled motions (e.g., foot-pedal trash cans) and direct generation of simplified collision primitives. Despite these limitations, the "perfect ground truth" and kinematic diversity provided by this toolkit represent a major step forward for generalizable robot learning.
Summary updated