New Receding Horizon Framework Enables Multi-Agent Deceptive Path Planning

Deceptive path planning lets autonomous agents hide their true intent from observers by deliberately deviating from optimal routes. Prior systems require full-horizon, end-to-end optimization that is computationally expensive, hard to adapt mid-mission, and rarely scales beyond a single agent. To solve this, researchers from Lehigh University and the Army Research Laboratory (Xubin Fang, Brian M. Sadler, Rick S. Blum) introduce a receding horizon multi-agent deceptive path planner built on a Boltzmann distribution. Instead of solving the entire path at once, the framework computes stochastic policies over short candidate trajectories within a rolling time window. The user defines a cost function balancing deception, resource use, and smoothness—plus optional coupling terms for coordination between multiple agents.

By iterating this cost locally, the system produces policies that dynamically trade off optimality against deception, all without any offline training. Deception levels and constraint adherence can be tuned in real time, so agents can instantly adjust as goals shift or obstacles appear. Simulation studies show the approach maintains deception while adapting to environmental changes, avoiding the costly recomputation of older end-to-end methods. The framework is intuitive to tune via a small handful of parameters, opening the door to new forms of dynamic, coordinated deception in robotics, defense, and autonomous swarms.