This paper is about a layered controller for a complex humanoid robot: namely, the iCub. We exploited a combination of precomputed models and machine learning owing to the principle of balancing the design effort with the complexity of data collection for learning. A first layer uses the iCub sensors to implement impedance control, on top of which we plan trajectories to reach for visually identified targets while avoiding the most obvious joint limits or self collision of the robot arm and body. Modeling errors or misestimation of parameters are compensated by machine learning in order to obtain accurate pointing and reaching movements. Motion segmentation is the main visual cue employed by the robot.
展开▼
