Using Bio-Realistic Gaussian-Shaped Population and Dopamine-Modulated STDP for Training a Self-Balancing System

Abstract

Human body balance is a gradual formation through repetition of actions, trial and error, and improving the mechanism of muscular-skeletal architecture for adapting to the demands of the environment. In the learning process, sensory receptors continuously send signals to the brain, then the brain to muscles and make a new signals pathway. Each time the body performs an action, millions of new synaptic connections are formed, and repetitive actions strengthen connections. So, a balanced body reuses the learned mechanism without performing any complex calculations. In contrast, the balance problem of a self-balancing robot has been solved by many different control algorithms. In this work, we propose a novel way to balance a two-wheeled self-balancing robot using bio-realistic Spiking Neural Networks (SNNs) to learn self-balancing, which is closely related to the way babies learn. To accomplish this, the gaussian shaped sensory neuronal population is connected with motor neurons through Spike-Timing-Dependent Plasticity (STDP) based synapses, further controlled with dopamine neurons. The key aspects of this approach are its bio-realistic nature and zero dependencies on data for adopting a new behavior compared to Deep Reinforcement Learning. Furthermore, this biologically-inspired mechanism can be used to improve the methodology for programming the robots to mimic Biological Intelligence.