Reproduce “ALRM: Agentic LLM for Robotic 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
Vitor Gaboardi dos Santos, Ibrahim Khadraoui, Ibrahim Farhat, Hamza Yous, Samy Teffahi, Hakim Hacid
The ALRM framework, developed by researchers at the Technology Innovation Institute, introduces a modular, agentic LLM architecture for robotic manipulation, enabling dynamic task planning and closed-loop execution. A new benchmark for multi-step tasks with diverse linguistic instructions was also established, revealing performance differences between large and small language models across execution modes.
Key insights
- Large-scale LLMs generally achieve higher success rates in Tool-as-Policy (TaP) mode due to better nested reasoning, but at the cost of increased latency, while Code-as-Policy (CaP) offers faster execution.
- Small-scale LLMs often struggle with TaP's multiple consecutive nested calls and high-level reasoning tasks, but some can achieve competitive CaP success rates with lower latency, such as Falcon-H1-7B.
- Large-scale LLMs maintain consistent performance across various linguistic categories, while small-scale models find high-level reasoning tasks to be the most challenging, indicating difficulties with inference from abstract instructions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.