Cooperative Gaussian process-based model predictive control for safe multi-agent navigation

Multi-agent systems, such as fleets of robots or drones, are increasingly deployed in logistics, inspection, and surveillance. These systems must reach their targets while maintaining safe separation, even under uncertain dynamics. This is challenging because unmodeled effects, disturbances, and sensor noise can degrade tracking performance and compromise safety. Model Predictive Control (MPC) is well suited for multi-agent navigation since it optimizes trajectories over a prediction horizon while enforcing input and state constraints. However, its performance depends on accurate models, and centralized formulations suffer from poor scalability and a single point of failure. We propose a cooperative Gaussian Process–augmented MPC (GP-MPC) framework that combines learning, chance-constrained safety, and distributed optimization. Each agent uses a Gaussian Process to learn its residual dynamics and quantify local uncertainty, incorporates this uncertainty into a chance-constrained collision-avoidance scheme, and coordinates only with neighbors through an ADMM-based distributed optimization method. This integration provides robustness to model errors and scalability to larger teams. The framework enables collision avoidance using only local uncertainty estimates, removing the need to share covariance information. By restricting computation and communication to each agent’s neighborhood, it maintains scalability and efficiency. Simulations show that the approach yields smoother and more efficient trajectories, faster convergence to targets, and reliable probabilistic safety compared to nominal and nonlinear MPC baselines. Convergence analysis further confirms robust consensus across a range of tuning parameters.