Gait Synthesis and Modulation for Quadruped Robot Locomotion Using a Simple Feed-Forward Network

Abstract

This paper describes a technique for statically stable gait synthesis for a quadruped robot using a simple Feed Forward Neural Networks (FFNN). A common approach for gait synthesis based on neural networks, is to use an implementation with Continuous Time Recurrent Neural Network (CTRNN) of arbitrary complex architecture as pattern generator for rhythmic limb motion. The preferred training method is implemented using genetic algorithms (GAs). However, to achieve the desired trajectory becomes an obstacle during the training process. This paper presents a much more simpler process converting a statically stable gait into actuator’s space via inverse kinematics; the training of the network is done with those references. By doing so, the training problem becomes a spatio-temporal machine learning problem. It is described a solution for trajectory generation combining a simple oscillator model with a Multilayer Feedforward Neural Network (MFNN) to generate the desired trajectory.