Orchestrating Method Ensembles to Adapt to Resource Requirements and Constraints during Robotic Task Execution

Kurzfassung

Robot behavior designers commonly select one method -- e.g. A* or RRT -- that is assumed to have the appropriate trade-off for a given domain between computational load, computation time, and the quality of the result of the method. We propose ensemble orchestration patterns, which evaluate multiple methods, and select the best result, thus exploiting the complementary advantages that alternative methods often have. By implementing different termination, preemption, constraint enforcement and selection schemes, different patterns lead to different (predictable) resource trade-offs. Thus, rather than selecting and committing to only one method, a designer chooses the appropriate pattern and constraints for the desired trade-off, and the pattern then realizes the selection on-line. We apply these patterns to various subtasks that are prevalent in our Surface Avatar ISS Technology Demonstration Mission, such as navigation, motion planning, and registration aswell as to a subtask in the service robotics domain in a simulated experiment. In our evaluation, we demonstrate that these patterns can effectively exploit increased resource budgets or relaxed constraints to find better solutions, and adapt the selection to different situations.