Speaker
Description
Distributed model predictive control (DMPC) is a powerful and general control technique to automate the behavior of a collaborating network of systems in a distributed manner. One key advantage is that, through the intuitive definition of constraints and the encoding of the collaborative goal in the cost function, a wide variety of collaborative tasks can be considered. Moreover, conceptually, DMPC can naturally deal also with nonlinear systems. In theory, this makes DMPC uniquely attractive in collaborative robotics and a potential enabler for self-reliant, robust robot swarms that can significantly outperform automation approaches relying on individual robots. In existing works, simulations and idealized experiments confirm these conceptual advantages of DMPC in challenging robotics scenarios. However, experiments under realistic conditions, with real robot-to-robot wireless communication, presently reveal limitations, as common communication networks can be overwhelmed by the sheer amount of communicated data. This can become a problem especially in faster robotics tasks that require decently high sampling rates. In particular, DMPC schemes communicate much more data than other distributed control techniques. Each controller sends at least one message of the size of an open-loop prediction per solution iteration, and one timestep may comprise multiple iterations. In consequence, communication is currently a key inhibitor preventing the widespread application of DMPC in real-world use cases in robotics.
This work presents viable approaches to overcome this limitation and tests them in formation control scenarios with mobile robots. Although the general methodology can be used for many applications and implementations of DMPC, the example application in robotics is particularly enticing due to many timely application areas that could benefit from reliable DMPC. The study considers nonholonomic mobile robots, which warrant a fully nonlinear treatment, bringing additional challenge to the task. To improve DMPC's communication, the contribution first analyzes what precisely is or can be communicated in a typical DMPC algorithm. Then, based on insights on the inherent structure of the communicated data, several ways to reduce the amount of communication are proposed and discussed, from simple, algorithmic ways to machine-learned approaches. Results from the most promising approaches reveal how far communication can be reduced while retaining the desired functionality of the distributed controller. There, a key motive to reduce communication is also to unlock the ability to use alternative communication methods other than traditional Wi-Fi, e.g., as a fallback to remain functional also when Wi-Fi communication is struggling or disturbed.
