How2comm: Communication-Efficient and Collaboration-Pragmatic Multi-Agent Perception Dingkang Yang 1,2 Kun Yang 1 Yuzheng Wang

As an emerging field, there are many challenges in building a robust multi-agent collaborative perception system. This paper addresses three of the biggest challenges, including communication redundancy, transmission delay, and collaboration heterogeneity, through tailored components. Although previous works have appropriately focused on these dilemmas, collaborative perception systems still have considerable room for improvements in the trade-off between perception performance and communication bandwidth. How2comm achieves superior performance with lower communication overheads in real V2V/X scenarios with limited communication capacity. We further discuss other realistic limitations and indicate future research directions to optimize our system. For the attack issue, How2comm could improve the robustness against potential adversarial attacks by introducing an adversarial training paradigm. In addition, by focusing on decoupled spatial regions and discovering perceptually critical information, How2comm is relatively less likely to be attacked. For the data bias issue, How2comm has demonstrated its preliminary generalization on real-world and simulated datasets with different configurations and sensor facilities.