Keywords

Abstract

Model Predictive Control (MPC) offers excellent constraint handling for multi-robot systems (MRS) but is highly sensitive to the choice of prediction horizon—a trade-off between performance and computational cost. This talk presents a Collective Reinforcement Learning framework that dynamically adjusts the prediction horizon of each robot based on the collective states of all robots. A versatile on-demand collision avoidance (VODCA) strategy further ensures safe operation under varying horizons. Numerical studies and real-world experiments on TurtleBot3 robots demonstrate scalability and robustness compared to fixed-horizon and existing variable-horizon MPC methods.