Reproduce “APEX: A Decoupled Memory-based Explorer for Asynchronous Aerial Object Goal 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
Daoxuan Zhang, Ping Chen, Xiaobo Xia, Xiu Su, Ruichen Zhen, Jianqiang Xiao, Shuo Yang
An asynchronous, decoupled memory-based explorer named APEX enables UAVs to autonomously navigate and locate target objects in complex aerial environments. This hierarchical agent achieves state-of-the-art performance on the UAV-ON benchmark, boosting Success Rate by 4.2% and Success-weighted Path Length by 2.8%, while significantly reducing operational latency and improving safe exploration.
Key insights
- Decoupling semantic reasoning, control generation, and target grounding into specialized, asynchronously operating modules effectively mitigates VLM latency and enables proactive, continuous UAV motion.
- Employing dynamic 3D grid-based maps allows for the robust integration of geometric and semantic information over time, providing a comprehensive spatial memory crucial for long-horizon aerial navigation.
- Training an RL-based action decision module with a composite reward function on map-based inputs translates high-level semantic understanding into stable and interpretable low-level control policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.