Reproduce “PLanAR: Planning-Language-Grounded Agentic Reasoning for Robot Manipulation” 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

Pengyuan Guo, Zhonghao Mai, Zhengtong Xu, Kaidi Zhang, Quan Khanh Luu, Heng Zhang, Zichen Miao, Arash Ajoudani and 3 more

PLanAR integrates Vision-Language Models with a Planning Domain Definition Language interface, enabling robots to perform complex, long-horizon manipulation tasks through structured, agentic reasoning. This framework utilizes closed-loop verification and replanning to achieve robust real-world performance, outperforming existing baselines, and revealing specific limitations of current VLMs in embodied reasoning.

Key insights

  • For robust robot agents, optimizing for closed-loop consistency and online verification is more critical than solely pursuing richer open-loop reasoning capabilities in VLMs.
  • Planning languages like PDDL provide a structured, verifiable interface that bridges powerful VLM understanding with the precise requirements of robot task planning, mitigating issues like VLM hallucination and improving consistency.
  • A compositional and modular agent architecture enables systematic analysis of VLM performance at each stage of a long-horizon task, facilitating targeted improvements and diagnosing specific bottlenecks in embodied reasoning.

Summary by alphaXiv.

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