Synaptic plasticity in a recurrent neural network for versatile and adaptive behaviors of a walking robot.

摘要: Walking animals, like insects, with little neural computing can effectively perform complex behaviors. For example, they walk around their environment, escape from corners/deadlocks, and avoid or climb over obstacles. While performing all these behaviors, also adapt movements to deal an unknown situation. As a consequence, successfully navigate through environment. The versatile adaptive abilities are the result of integration several ingredients embedded in sensorimotor loop. Biological studies reveal that include dynamics, plasticity, sensory feedback, biomechanics. Generating such behaviors for many degrees-of-freedom (DOFs) walking robot is challenging task. Thus, this study, we present bio-inspired approach solve Specifically, combines mechanisms exteroceptive consist processing modular locomotion control. based on small recurrent network consisting two fully connected neurons. Online correlation-based learning synaptic scaling applied adequately change connections network. By doing so, exploit dynamics (i.e., hysteresis effects single attractors) generate different turning angles short-term memory robot. information transmitted as descending steering signals control which translates into motor actions. result, its angle avoiding obstacles situations. adaptation enables sharp corners deadlocks. Using backbone joint allows Consequently, it explore environments. We firstly tested our physical simulation environment then real biomechanical AMOSII 19 DOFs adaptively world.