Robot Locomotion Control Using Central Pattern Generator with Non-linear Bio-mimetic Neurons

Abstract

Central pattern generators (CPG) generate rhythmic gait patterns that can be tuned to exhibit various locomotion behaviors like walking, trotting, etc. CPGs inspired by biology have been implemented previously in robotics to generate periodic motion patterns. This paper aims to take the inspiration even further to present a novel methodology to control movement of a four-legged robot using a non-linear bio-mimetic neuron model. In contrast to using regular leaky integrate and fire (LIF) neurons to create coupled neural networks, our design uses non-linear neurons constituting a mixed-feedback (positive and negative) control system operating at multiple timescales (fast, slow and ultraslow ranging from sub-ms to seconds), to generate a variety of spike patterns that control the robotic limbs and hence its gait. The use of spikes as motor control signals allows for low memory usage and low latency operation of the robot. Unlike LIF neurons, the bio-mimetic neurons are also jitter tolerant making the CPG network more resilient and robust to perturbations in the input stimulus. As a proof of concept, we implemented our model on the Petoi Bittle bot, a quadruped pet dog robot and were able to reliably observe different modes of locomotion-walk, trot and jump. Four bio-mimetic neurons forming a CPG network to control the four limbs were implemented on Arduino microcontroller and compared to a similar CPG built using four LIF neurons. The differential equations for both neurons were solved real-time on Arduino and profiled for memory usage, latency and jitter tolerance. The CPG using bio-mimetic non-linear neurons used marginally higher memory (378 bytes, 18% higher than LIF neurons), incurred insignificant latency of 3.54ms compared to motor activation delay of 200ms, while providing upto 5-10x higher jitter tolerance.