Robot control based on constrained optimization is a powerful method to control complex robotic tasks based on constraints. The idea is to break down the overall control problem into simultaneously running subtasks and merge them into a coherent robot control signal using constrained optmization. Despite on its capabilities in specifying complex robot control problems, the control solution is usually governed by a rather large set of parameters, which have to be tuned manually. Combining constraint-based robot control with learning approaches might not only simplify the cumbersome process of parameter selection, but also provide novel, context-adaptive control solutions.
In this talk I would like to illustrate my recent progress on context-adaptive robot control based on constraints. For this purpose, I will present an approach for learning task constraints in a real-time human-robot collision avoidance scenario. Experimental results are presented on a dual-arm industrial robot system, which is supposed to avoid collisions with a human operator entering its workspace.