Reaching movement generation with a recurrent neural network based on learning inverse kinematics for the humanoid robot iCub

We present a dynamical system approach that couples task and joint space by means of an attractor-based content addressable memory. The respective recurrent reservoir network simultaneously provides a novel control framework for goal directed movement generation. The network first learns to associate end effector coordinates with joint angles by means of reservoir attractor states and thereby implements forward and inverse kinematics. Generalization of the learned kinematics to a wide range of untrained target positions is achieved by modulating the attractors with desired input states. We show that this representation of the static kinematic mapping within a dynamical system also enables smooth trajectory generation by exploiting the transient network dynamics when approaching an attractor state. A further strength of the proposed approach is that efficient online learning and execution of the network makes it real-time capable. We demonstrate the network´s generalization abilities and evaluate controller properties systematically for arm movements of the humanoid robot iCub.