Reproduce “Asking for Help: Failure Prediction in Behavioral Cloning through Value Approximation” 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

Cem Gokmen, Daniel Ho, Mohi Khansari

Researchers at Everyday Robots and Stanford University developed Behavioral Cloning Value Approximation (BCVA), a framework that jointly learns a policy and a state value function to predict impending robot failures. This approach demonstrated superior performance in real-world mobile manipulation, achieving an 83% F1 score for failure prediction in a latched-door opening task by accounting for environmental dynamics through pixel differences.

Key insights

  • Integrating a state value function directly into a Behavioral Cloning framework enables data-efficient, preemptive failure prediction without the computational overhead of full Reinforcement Learning.
  • Utilizing failure examples from on-policy rollouts, in addition to expert demonstrations, significantly enhances the learned state representation and the accuracy of the value function.
  • Customizing the distance metric for calculating discounted returns (e.g., using pixel-level changes) can more effectively capture task dynamics and improve value approximation in vision-based mobile manipulation.

Summary by alphaXiv.

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