SurveilNav uses 206 cameras across 74 floors to revolutionize robot navigation

Researchers led by Ming-Ming Yu have introduced SurveilNav, a novel collaborative navigation framework that bridges the gap between mobile robots and fixed surveillance cameras. Published on arXiv and accepted at ICRA 2026, the system leverages a massive indoor dataset built on Habitat-Sim featuring 206 cameras across 74 floors. This setup enables agents to exploit multi-view surveillance information, addressing the core limitation of single-robot perception: restricted field of view and inherent blind spots of static cameras. SurveilNav combines active camera scheduling, joint 2D/3D mapping, vision-language model (VLM) based value estimation, and collaborative target verification to synergize the robot's dynamic local perception with the surveillance system's global static view.

The framework's key innovation lies in orchestrating which cameras to activate and when, then fusing that data with the robot's on-board sensors to build a unified spatial understanding. The VLM component estimates the value of unexplored regions, guiding the robot toward promising areas while avoiding redundant coverage. On the challenging HM3D dataset, SurveilNav substantially outperformed existing methods in both exploration efficiency and navigation success rate. The authors highlight strong potential for applications in large-scale search operations, home environments, and rescue missions where a mix of fixed infrastructure and mobile agents is already available. By turning every camera into a collaborative teammate, SurveilNav points toward a future where robots no longer navigate alone.