Reproduce “Learning Robust Execution in Robotic Manipulation with Agentic Reinforcement Learning” 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

Xiaopeng Zhang, Yueyang Weng, Qi Liu, Yongjin Mu, Yanjie Li

An agentic reinforcement learning framework improves the robustness of robotic manipulation policies by learning high-level decisions for execution management and recovery. This approach, developed by researchers including those from Harbin Institute of Technology Shenzhen, enhanced task success rates by up to 39.2% under kinematic disturbances without requiring modifications to the underlying low-level policies.

Key insights

  • Monitoring execution quality through combined local and global metrics provides effective real-time signals for identifying degradation, enabling the high-level agent to make informed recovery decisions.
  • A lightweight agentic policy can be trained to adaptively select appropriate recovery actions, proportionate to the type and severity of execution degradation, effectively managing complex low-level policies without requiring their modification or retraining.
  • Decoupling execution management from low-level action generation offers a cost-effective and generalizable approach to enhance robustness across various learning-based manipulation policies and diverse tasks.

Summary by alphaXiv.

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