Privacy-Preserving MAPF Keeps Agent Paths Hidden During Planning and Execution

In a new paper accepted at AAMAS 2026, researchers from Ben-Gurion University address a critical gap in multi-agent path finding (MAPF): privacy. Standard MAPF assumes agents can share planned paths to avoid collisions, but in applications like warehouse robotics or autonomous vehicle coordination, agents may not want to reveal their locations. The authors formulate two distinct privacy constraints: planning-level privacy (hiding intended paths during coordination) and execution-level privacy (preventing agents from sensing each other’s positions during movement).

For planning-level privacy, they propose adding dummy agents to the planning process to obscure true paths without compromising collision avoidance. For execution-level privacy, they adapt two widely-used algorithms—PIBT (Priority Inheritance with Backtracking) and LaCAM (Labeled Conflict Avoidance with Multi-agent)—to run with limited sensing information. A novel post-processing technique further optimizes the total path cost without sacrificing privacy. The 16-page paper includes extensive empirical evaluations showing that post-processing significantly reduces solution costs. This work opens the door to deploying MAPF in privacy-sensitive environments like logistics, defense, and autonomous transportation.