Reproduce “When2com: Multi-Agent Perception via Communication Graph Grouping” 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
Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, Zsolt Kira
When2com proposes a learning-based framework that enables multi-agent systems to dynamically determine when and with whom to communicate for collaborative perception tasks. This approach significantly reduces communication bandwidth while enhancing accuracy, demonstrating improvements in semantic segmentation and 3D shape recognition by selectively leveraging relevant peer information.
Key insights
- Intelligent, resource-efficient communication decisions (both 'when' and 'who' to communicate with) can be learned end-to-end, driven solely by the downstream perception task's performance.
- An asymmetric query-key mechanism allows agents to broadcast low-bandwidth queries while maintaining high-dimensional internal representations for effective matching and integration.
- A self-matching score (self-attention) is crucial for an agent to determine if its local observations are sufficient, enabling it to avoid unnecessary communication and prevent the intake of detrimental information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.