Learning-Based Auto-Tuning for Motion Controllers of Mobile Robots

Abstract: An auto-tuner of the parameters of a mobile robots motion controlleris developed to improve its performance. The generality of the auto-tunerallows for similar applications on other robots or controllers. The performanceis defined by a proprietary objective function for the purpose ofreducing positional error, wear and energy consumption while retainingstability. The function evaluates data recorded during simulated runs ofthe robot. Sequential model-based Bayesian optimization (SMBO) is evaluatedwith different design choices. Classic black box optimizers, namelygrid search, random search and Latin hypercube sampling are used asbenchmarks, as well as manual tuning. It is found that SMBO performson par with manual tuning after less than 1000 iterations searching avery large feature space, making use of no prior knowledge for settingthe boundaries. When instead guided by relatively narrow boundariesaround the manually chosen parameter values less than 5 iterations areneeded and the performance continues improving for another 600 iterations.The SMBO demands much fewer evaluations than grid search; itoutperformed grid search even when running 20 times more iterations.However, the difference in performance of different design choices for theSMBO was negligible.