A dynamical system for online learning of periodic movements of unknown waveform and frequency

The paper presents a two-layered system for learning and encoding a periodic signal onto a limit cycle without any knowledge on the waveform and the frequency of the signal, and without any signal processing. The first dynamical system is responsible for extracting the main frequency of the input signal. It is based on adaptive frequency phase oscillators in a feedback structure, enabling us to extract separate frequency components without any signal processing, as all of the processing is embedded in the dynamics of the system itself. The second dynamical system is responsible for learning of the waveform. It has a built-in learning algorithm based on locally weighted regression, which adjusts the weights according to the amplitude of the input signal. By combining the output of the first system with the input of the second system we can rapidly teach new trajectories to robots. The systems works online for any periodic signal and can be applied in parallel to multiple dimensions. Furthermore, it can adapt to changes in frequency and shape, e.g. to non-stationary signals, and is computationally inexpensive. Results using simulated and hand-generated input signals, along with applying the algorithm to a HOAP-2 humanoid robot are presented.

[1]  Jun Nakanishi,et al.  Learning Attractor Landscapes for Learning Motor Primitives , 2002, NIPS.

[2]  A. Ijspeert,et al.  Dynamic hebbian learning in adaptive frequency oscillators , 2006 .

[3]  Ludovic Righetti,et al.  Engineering entrainment and adaptation in limit cycle systems , 2006, Biological Cybernetics.

[4]  Andrej Gams,et al.  Imitating human acceleration of a gyroscopic device , 2007, Robotica.

[5]  Jun Nakanishi,et al.  Movement imitation with nonlinear dynamical systems in humanoid robots , 2002, Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No.02CH37292).

[6]  Christopher G. Atkeson,et al.  Constructive Incremental Learning from Only Local Information , 1998, Neural Computation.

[7]  Barak A. Pearlmutter.  Gradient calculations for dynamic recurrent neural networks: a survey , 1995, IEEE Trans. Neural Networks.

[8]  Leon Zcaron,et al.  Robotic yo-yo: modelling and control strategies , 2006 .

[9]  Matthew M. Williamson,et al.  Neural control of rhythmic arm movements , 1998, Neural Networks.

[10]  Jun Nakanishi,et al.  Learning rhythmic movements by demonstration using nonlinear oscillators , 2002, IEEE/RSJ International Conference on Intelligent Robots and Systems.

[11]  Ludovic Righetti,et al.  Programmable central pattern generators: an application to biped locomotion control , 2006, Proceedings 2006 IEEE International Conference on Robotics and Automation, 2006. ICRA 2006..

[12]  Jun Nakanishi,et al.  Nonlinear dynamical systems for imitation with humanoid robots , 2001 .

[13]  Yoshihiko Nakamura,et al.  Polynomial design of the nonlinear dynamics for the brain-like information processing of whole body motion , 2002, Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No.02CH37292).

[14]  Leon Zlajpah.  Robotic yo-yo: modelling and control strategies , 2006, Robotica.