Reproduce “MaP-AVR: A Meta-Action Planner for Agents Leveraging Vision Language Models and Retrieval-Augmented Generation” 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
Zhenglong Guo, Yiming Zhao, Feng Jiang, Heng Jin, Zongbao Feng, Jianbin Zhou, Siyuan Xu
Li Auto researchers introduced MaP-AVR, a robotic task planner utilizing Vision-Language Models and Retrieval-Augmented Generation to generate plans using robot-intrinsic 'meta-actions' instead of human-centric skills. This system demonstrated a 43.13% task success rate in simulation, outperforming a baseline method's 13.75%.
Key insights
- Abstracting robot skills into 'meta-actions' based on intrinsic robot functionalities offers greater generality and direct executability than traditional human-centric skill definitions, enhancing task versatility.
- The integration of Retrieval-Augmented Generation (RAG) significantly improves a VLM's ability to produce consistent, correctly formatted, and rational task plans by providing relevant in-context learning examples.
- A self-augmenting database for RAG enables the planning system to continuously learn from successfully executed tasks, progressively enhancing its robustness and expanding its knowledge base for future scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.