Autogeneration of Mission-Oriented Robot Controllers Using Bayesian-Based Koopman Operator

Model-based robot controllers require customized control-oriented models, involving expert knowledge and trial and error. Remarkably, the Koopman operator enables the control-oriented model identification through the input–output mapping set, breaking through the barriers of the customization services. However, in recent years, research on Koopman-based robot control has mostly focused on lifting function construction, deviating from the original intention of improving the controller performance. Thus, we propose a robot controller autogeneration framework using the Bayesian-based Koopman operator, significantly releasing labor and eliminating the design obstacle. First, we introduce the Koopman-based system identification method and offer the basic lifting function design criteria. Then, a Bayesian-based optimization strategy with resource allocation is designed, which allows for the simultaneous optimization of the lifting function and the controller. Next, taking model-predictive control (MPC) as an example, a mission-oriented controller autogeneration framework is developed. Simulation and experimental results indicate that, under various robots and data sources, the proposed framework can effectively generate the robot controllers and perform with a far greater level of mission accuracy than the unoptimized Koopman-based MPC. Meanwhile, the proposed technique exhibits an obvious compensation effect against disturbances, demonstrating its practicability in robot control.