Correcting for changes: expected perception-based control for reaching a moving target

摘要

Expected perception (EP)-based control systems use the robotic system?s internal models and interaction with the environment to predict the future response of their sensory inputs. By comparing the sensory predictions with actual sensory data, the EP control system monitors the error between the predicted and the actual sensor observations. If the error is small, the system may decide to neglect the input and skip any corrective action, thus saving computational and energy resources. If the mismatch is large, the system will further process the sensor signal to compute a corrective action through feedback. So far, EP systems have been implemented for predictions based on a robot?s motion. In this article, an EP system is applied to predict the dynamics and anticipate the motion of an external object. The new control system is implemented in a humanoid robot, the iCub. The robot reaches in anticipation for an object?s future position by predicting its trajectory and correcting the arm?s position only when necessary. The results of the EP-based controller are analyzed and compared against a standard controller. The new EP-based controller is less computationally demanding and more energy efficient for a marginal loss in the tracking error.