Reproduce “ReSteer: Quantifying and Refining the Steerability of Multitask Robot Policies” 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

Zhenyang Chen, Alan Tian, Liquan Wang, Benjamin Joffe, Yingyan Celine Lin, Yuxiao Chen, Siddharth Karamcheti, Danfei Xu

ReSteer introduces a framework to quantify and enhance the ability of multitask robot policies to adapt to new instructions mid-execution, showing a 2.2x improvement in real-world steering success while maintaining high single-task proficiency.

Key insights

  • Conditional Mutual Information (CMI) serves as an effective and efficient proxy for steerability, showing a strong correlation (r=0.74) with ground-truth task-switching success and enabling diagnosis of instruction-blind regions.
  • The limited steerability of current multitask robot policies is largely attributable to a data distribution pathology where static language in training data causes policies to prioritize state over dynamic language conditioning.
  • Targeted data augmentation that includes explicit mid-execution instruction switches, especially when guided by CMI to prioritize low-steerability states, significantly improves sample efficiency and policy responsiveness.

Summary by alphaXiv.

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