Reproduce “Learning to Look: Seeking Information for Decision Making via Policy Factorization” 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
Shivin Dass, Jiaheng Hu, Ben Abbatematteo, Peter Stone, Roberto Martín-Martín
Researchers from The University of Texas at Austin introduce DISaM, a dual-policy framework enabling robots to learn to actively seek task-relevant information and then use it for manipulation. This method, leveraging an intrinsic reward for information-seeking, consistently achieved higher success rates than baseline approaches across various complex simulated and real-world tasks.
Key insights
- Factorizing the problem into distinct information-seeking and manipulation policies simplifies learning and addresses the challenges of sparse rewards in long-horizon, partially observable tasks.
- An intrinsic reward signal, generated by comparing the manipulation policy's actions based on inferred versus ground-truth context, provides a dense and effective learning signal for the information-seeking agent.
- Dynamically switching between information-seeking and manipulation actions based on the manipulation policy's uncertainty enables efficient and adaptive task execution, allowing the robot to gather information only when needed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.