Adaptive Manipulator Control using Active Inference with Precision Learning

Abstract

Active inference provides a framework for decision- making where the optimization is achieved by minimizing free- energy. Previous work has used this framework for control and state-estimation of a robotic manipulator. This required manual definition of precision matrices which serve as controller gains. This paper provides an implementation for control and state-estimation where the precision matrices are tuned during execution-time (precision learning). Learning the precision ma- trices means automatically adjusting the controller’s gains which decreases oscillations and overshoot.