Movement Control of Tensegrity Robot

Certain kinds of mollusks can perform flexible actions by changing body shape. Octopi, in particular, have little in the way of a skeletal structure, and their body frame consists mostly of muscular fibers roughly equivalent to actuators. Here, we have applied tensegrity structure[3,1,2] as a flexible body structure, just as the muscular fibers of an octopus. This tensegrity structure is used in a robot that moves by changing form, creating various shapes in the process. Since the most effective means of movement in the case of a simple body is rotation, sensor information fluctuates drastically. Here, we assess control of such a robot by rule-based general reinforcement learning and by a recurrent neural network using neural oscillators. When reinforcement learning was used, because the sensor information is prone to drastic change, the robot had only a small number of chances to learn in each state, making effective control regardless of the state difficult. However, a certain level of control was possible during conditions in which the sensor information did not alter as significantly. When neural oscillators were used, a considerable amount of time was required for updating link weights, but smooth movement was achieved because of the rhythm generated by the oscillators.