Reproduce “AOMGen: Photoreal, Physics-Consistent Demonstration Generation for Articulated Object Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yulu Wu, Jiujun Cheng, Haowen Wang, Dengyang Suo, Pei Ren, Qichao Mao, Shangce Gao, Yakun Huang
AOMGen synthesizes photorealistic and physically consistent demonstration data for articulated object manipulation from minimal real-world inputs, addressing data scarcity for robot learning. The framework achieved an average 98% success rate in simulator replays of generated trajectories and improved VLA model performance on unseen objects from 15% to 65% success rate when fine-tuned with diverse AOMGen data.
Key insights
- Synthesizing photorealistic (via 3DGS) and physically consistent (via adapted interactions) data is essential for scalable data generation that can effectively bridge the sim-to-real gap for complex robot manipulation tasks.
- Accurate recovery of articulated object motion and joint parameters from limited visual input, coupled with contact point detection, enables physically plausible interaction adaptation for new objects.
- Generating diverse data through category-level object replacement and arbitrary pose generalization from a single real-world demonstration allows for significant improvement in VLA model robustness and generalization to unseen scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.