摘要：AbstractMany systems exhibit multiple modes of operation, where the real mode is not known a priori. Examples are the grasping of objects by a robot, where a variety of object forms and stiffnesses are possible, or a quadcopter lifting and dropping several objects with different weights. In such cases it is challenging to design a controller that ensures robustness and safety for all possible modes without jeopardising performance. To this end, we present a learning supported predictive control approach for multi-mode uncertain environments with robustness guarantees. The approach allows to fuse a priori knowledge via including multiple models of the system with on-line and off-line learning to improve the system models and thus the overall performance. Guaranteed robustness is ensured by decoupling it from performance. While learned and improved models are used for performance optimization, robustness is guaranteed by ensuring that all possible modes respect the constraints and are repeatedly feasible. The ideas are confirmed considering an UAV package delivery example.
关键词：KeywordsNonlinear model predictive controlmulti-mode systemsmachine learningrobustness
