Reproduce “MoPA: Coordinated Mobile Manipulation via Subsystem-Specific Perception Alignment” 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

Guangyu Chen, Qiwei Liang, Shaolong Zhu, Tianxing Chen, Zikuan Xiao, Yifan Xie, Lingfeng Zhang, Ping Luo and 2 more

MoPA separates visual queries for mobile and manipulation control, matching each query stream to its action stream while allowing the action streams to communicate and update together. On ManiSkill-HAB, it has the highest macro-averaged success across three suites, including 88.3% on SetTable. Across four real-world tasks, one jointly trained policy averages 76.3% success, 12.5 percentage points above π0.5, but ties it on microwave retrieval.

Key insights

  • Explicit alignment between a perceptual stream and its corresponding action branch is more effective than simply exposing both branches to a larger shared representation.
  • Perceptual queries that evolve alongside intermediate actions can adapt scene understanding to the current control trajectory rather than serving only as static conditioning.
  • Navigation-relevant queries attend more to floors and broad scene structure, whereas manipulation queries focus on target-object regions, providing evidence that the streams learn distinct roles.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.