Reproduce “IndoorUAV: Benchmarking Vision-Language UAV Navigation in Continuous Indoor Environments” 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

Xu Liu, Yu Liu, Hanshuo Qiu, Yang Qirong, Zhouhui Lian

Researchers at Peking University developed IndoorUAV, a benchmark for Vision-Language Navigation (VLN) with Unmanned Aerial Vehicles (UAVs) in continuous indoor environments, featuring over 50,000 instruction-trajectory pairs across 1,000 3D scenes. The proposed hierarchical IndoorUAV-Agent, leveraging GPT-4o for task decomposition and a fine-tuned policy for execution, achieved a 7.29% success rate in seen environments for long-horizon tasks, outperforming existing baselines.

Key insights

  • Indoor aerial VLN requires simultaneous advancements in high-level language understanding for multi-step goals and low-level, precise 3D aerial control.
  • Task decomposition, leveraging large language models like GPT-4o, is crucial for improving performance on complex long-horizon VLN tasks by simplifying multi-step instructions.
  • The unconstrained 3D movement and cluttered nature of indoor environments necessitate fine-grained, continuous control over 4 Degrees-of-Freedom (DoF) for UAVs.

Summary by alphaXiv.

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