Reproduce “Warehouse Spatial Question Answering with LLM Agent” 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
Hsiang-Wei Huang, Jen-Hao Cheng, Kuang-Ming Chen, Cheng-Yen Yang, Bahaa Alattar, Yi-Ru Lin, Pyongkun Kim, Sangwon Kim and 3 more
The SpatialAgent framework, an LLM agent system developed by a collaborative research group, secured first place in the 9th AI City Challenge Track 3 by integrating Gemini 2.5-Flash with custom-trained lightweight perception models for warehouse spatial question answering. This data-efficient approach achieved 95.86% accuracy on the Physical AI Spatial Intelligence Warehouse Benchmark.
Key insights
- Complex spatial reasoning in visual environments can be achieved data-efficiently by augmenting a general-purpose LLM with a collection of lightweight, specialized perception models rather than extensive MLLM finetuning.
- A modular LLM agent architecture that leverages function-calling capabilities allows for flexible expansion of specific spatial competencies and robust, iterative reasoning for complex queries.
- Effectively bridging raw visual perception (via object masks) with the LLM's high-level cognitive reasoning enables nuanced understanding and interpretation of intricate spatial relationships in real-world scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.