Reproduce “GROVE: Grounded Pedestrian Simulation via Natural Language for Interactive Social Robot Navigation” 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
Duc Tai Nguyen, Volodymyr Shcherbyna, Anh Do Duc, Zhengcheng Shen, Teham Buiyan, Linh Kästner
GROVE is a text-to-scenario framework for pedestrian simulation that generates realistic and socially challenging environments for training social robots, integrating natural language understanding with high-fidelity behavioral and geometric planning. It consistently produced semantically coherent, structurally realistic, and geometrically correct simulations, achieving the highest VLM-based evaluation scores for prompt alignment and plausibility across diverse scenarios while reducing token count by 35-49% using optimized presets.
Key insights
- Separating semantic reasoning, behavioral planning via behavior trees, and geometric motion planning allows for robust, interpretable, and modular crowd simulation from natural language.
- Retrieval-Augmented Generation (RAG) effectively grounds LLM-based behavior synthesis by injecting task-relevant node specifications, mitigating context length issues and hallucination.
- Combining symbolic behavior trees with explicit global path planning for static obstacles ensures both semantically consistent intent and geometrically plausible, collision-free navigation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.