Reproduce “Opening Articulated Structures in the Real World” 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
Arjun Gupta, Michelle Zhang, Rishik Sathua, Saurabh Gupta
A modular mobile manipulation system (MOSART) developed at the University of Illinois at Urbana-Champaign achieves a 61% success rate in opening 31 novel articulated objects across 13 diverse real-world environments. The system demonstrates that modular approaches can significantly outperform state-of-the-art end-to-end imitation learning methods for zero-shot generalization in complex physical interaction tasks.
Key insights
- Modular systems, like MOSART, can outperform state-of-the-art end-to-end imitation learning methods in generalizing to novel real-world mobile manipulation tasks.
- The primary bottleneck for end-to-end mobile manipulation in unstructured environments is robust perception, accounting for 59% of system failures, rather than precise control.
- Articulation parameter estimation models developed in isolation often struggle with noisy, robot-centric viewpoints encountered in real-world deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.