How to Harness the Dynamics of Soft Body: Timing Based Control of a Simulated Octopus Arm via Recurrent Neural Networks

References [1] D. Trivedi, C. D. Rahn, W. M. Kier and I. D. Walker, “Soft robotics: Biological inspiration, state of the art, and future research,” Applied Bionics and Biomechanics, vol. 5 (3), 2008, pp 99-117. [2] G. Sumbre, Y. Gutfreund, G. Fiorito, T. Flash and B. Hochner, “Control of Octopus Arm Extension by a Peripheral Motor Program,” Science, vol. 293, 2001, pp 1845–1848. [3] Y. Yekutieli, R. Sagiv-Zohar, R. Aharonov, Y. Engel, B. Hochner and T. Flash, “Dynamic Model of the Octopus Arm. I. Biomechanics of the Octopus Reaching Movement,” J. Neurophysiol., vol. 94, 2005, pp 1443-1458. [4] R. Pfeifer and C. Scheier, Understanding Intelligence, Cambridge, MA: MIT Press; 1999. Timing-Based Control via Recurrent Neural Network Control Architecture: The control architecture is based on a recurrent neural network (RNN), in combination with a feed forward network (FFN). The main body of the RNN controls the angle of the arm base and the timing to send a signal to the low-level control (PNS). The accompanying network decides the power of the signal and the angle hat is required to achieve the reaching behavior. Task and training of the network: In order to determine the performance of the networks, we established the reaching tasks by using a simulated octopus arm. As revealed in octopus biology, the octopus starts to create a bend on the dorsal side of its arm and, through the bend propagation, its arm approaches the object from the ventral side. The important point here is the time it takes for the bend to form. Our aim is to autonomously control the time lag in the network. In order to achieve the reaching behavior toward the object, the networks are required to exploit the physical dynamics of the arm. For the learning of the networks we applied an incremental learning strategy. Unlike the conventional supervised learning case, we do not predefine the learning sets, but rather collect the learning sets by actually running the arm.