Reproduce “Know Where You're Uncertain When Planning with Multimodal Foundation Models: A Formal Framework” 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

Neel P. Bhatt, Yunhao Yang, Rohan Siva, Daniel Milan, Ufuk Topcu, Zhangyang Wang

A formal framework is presented for disentangling and quantifying perception and decision uncertainty in multimodal foundation models used for robotic planning. The approach, developed by researchers at The University of Texas at Austin, leads to up to a 5% higher probability of satisfying task specifications and reduces plan violation in real-world driving scenes through targeted interventions like active sensing and automated refinement.

Key insights

  • Uncertainty in multimodal foundation models can be formally disentangled into distinct perception and decision components, enabling targeted analysis and intervention.
  • Integrating formal methods with machine learning allows for objective quantification of decision uncertainty by verifying plan compliance against temporal logic specifications, reducing reliance on human annotations.
  • Targeted interventions, such as active sensing for perception uncertainty and automated refinement for decision uncertainty, lead to more reliable and robust robotic planning.

Summary by alphaXiv.

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