Infinigen-Sim: Procedural Generation of Articulated Simulation Assets

Abhishek Joshi1, Beining Han1, Jack Nugent1, Yiming Zuo1, Jonathan Liu1, Hongyu Wen1, Stamatis Alexandropoulos1, Tao Sun1,2, Alexander Raistrick1, Gaowen Liu3, Yi Shao2, Jia Deng1,
1Princeton University, 2McGill University, 3Cisco
arXiv Code

Generate unlimited articulated objects compatible with Isaac, MuJoCo, ManiSkill3, and many other simulation frameworks!

Getting Started
Articulated door 1 Articulated door 2 Articulated door 3 Articulated door 4 Articulated door 5 Articulated door 6
Articulated fridge 1 Articulated fridge 3 Articulated fridge 4 Articulated fridge 2 Articulated fridge 5 Articulated fridge 6
Articulated toaster 1 Articulated toaster 3 Articulated toaster 4 Articulated toaster 2 Articulated toaster 5 Articulated toaster 6

Abstract

We introduce Infinigen-Sim, a toolkit which enables users to create diverse and realistic articulated object procedural generators. These tools are composed of high-level utilities for use creating articulated assets in Blender, as well as an export pipeline to integrate the resulting assets into common robotics simulators. We demonstrate our system by creating procedural generators for 5 common articulated object categories. Experiments show that assets sampled from these generators are useful for movable object segmentation, training generalizable reinforcement learning policies, and sim-to-real transfer of imitation learning policies.

Provided Articulated Asset Categories

We've provided procedural generators for 5 articulated asset categories: doors, dishwashers, lamps, refrigerators, and toasters.
Articulated door 7 Articulated dishwasher 1 Articulated lamp 1 Articulated fridge 7 Articulated toaster 7

Key Advantages

  1. Accurate Joints and Highly Detailed Assets: Assets in Infinigen-Sim are procedurally generated, giving us full control over every part and joint. This ensures correct joint configurations by design. Infinigen-Sim also supports high-detail modeling, enabling the inclusion of small, articulated components.
  2. Unlimited Kinematic Variation: Users can define custom parameter distributions to vary asset properties that affect downstream performance. For example, the distance between a door handle and the door hinge influences a robot's trajectory. By adjusting this parameter, users can generate a diverse set of door configurations that promote more robust policies.
  3. Combinatorial Coverage: Many asset parameters can be varied independently, allowing Infinigen-Sim to generate a wide and diverse range of asset configurations. This independence enables combinatorial diversity, making it possible to systematically explore how individual factors influence downstream performance in robot learning tasks.

Creating Your Own Procedural Articulated Assets

Infinigen-Sim extends Blender Geometry and Shader Nodes, a popular artist-friendly system designed for procedural generation. This system allows you to model meshes and materials by composing primitives, geometric transformations, and arithmetic operations represented as nodes in a directed graph. Assets can be made in three easy steps:

  1. Create your articulated asset in Blender using our custom joint nodes.
  2. Define custom parameter distributions.
  3. Export to your simulation format of choice using Infinigen-Sim.
More information on how to create your own custom assets coming soon!

Blender demo

Successful Sim-to-Real

Through Infinigen-Sim, we see a significant boost in performance for opening doors in the real world. The task for the robot is to first turn the handle and then push the door open. We train an ACT (Zhao et al., 2023) policy using trajectories collected using a motion planner in simulation. We saw that a policy trained using purely Infinigen-Sim assets achieved a success rate of 70% in the real world while a policy trained on baseline assets from PartNet-Mobility (Xiang et al., 2020) achieved a 0% success rate.

Real demo 2 Real demo 1 Real demo 3

BibTeX

@misc{joshi2025infinigensimproceduralgenerationarticulated,
      title={Infinigen-Sim: Procedural Generation of Articulated Simulation Assets}, 
      author={Abhishek Joshi and Beining Han and Jack Nugent and Yiming Zuo and Jonathan Liu and Hongyu Wen and Stamatis Alexandropoulos and Tao Sun and Alexander Raistrick and Gaowen Liu and Yi Shao and Jia Deng},
      year={2025},
      eprint={2505.10755},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2505.10755}, 
}

Website design inspired by Nerfies.