Theoretical guarantees for predictive control in adaptive multi-agent scenarios

Motivation

Bridging the gap between theoretic research in distributed and learning-based model predictive control and multi-agent applications

Goals

• Modularity – local modification of robotic networks, including learning of unmodeled parts of the system dynamics based on communicated and sensed data (learning from neighbors and the environment)
• Decoupling design processes of the local controllers and the global coordination scheme
• Taking into account the specific dynamic properties of practical real-world robots
• Show practical applicability in realistic simulative and experimental scenarios, e.g. in robotics applications

Methods

• combination of distributed and learning-based model predictive control
• Gaussian process (GP) regression to infer unmodeled parts of the system dynamics at runtime
• continued development of distributed control hardware test benches, producing real-world data