Learning to Balance While Reaching: A Cerebellar-Based Control Architecture for a Self-balancing Robot

In nature, Anticipatory Postural Adjustments (APAs) are actions that precede predictable disturbances with the goal of maintaining a stable body posture. Neither the structure of the computations that enable APAs are known nor adaptive APAs have been exploited in robot control. Here we propose a computational architecture for the acquisition of adaptive APAs based on current theories about the involvement of the cerebellum in predictive motor control. The architecture is applied to a simulated self-balancing robot (SBR) mounting a moveable arm, whose actuation induces a perturbation of the robot balance that can be counteracted by an APA. The architecture comprises both reactive (feedback) and anticipatory-adaptive (feed-forward) layers. The reactive layer consists of a cascade-PID controller and the adaptive one includes cerebellar-based modules that supply the feedback layer with predictive signals. We show that such architecture succeeds in acquiring functional APAs, thus demonstrating in a simulated robot an adaptive control strategy for the cancellation of a self-induced disturbance grounded in animal motor control. These results also provide a hypothesis for the implementation of APAs in nature that could inform further experimental research.