Two-Legged Robot Motion Control With Recurrent Neural Networks

Legged locomotion is a desirable ability for robotic systems thanks to its agile mobility and wide range of motions that it provides. In this paper, the use neural network-based nonlinear controller structures which consist recurrent feedforward layers have been examined in dynamically stable walking problem two-legged robots. detail, hybrid controllers, long short-term memory type neuron models employed at layers, are utilized feedback paths. To train these networks, supervised learning data sets created by using biped robot platform controlled central pattern generator. Then, networks perform gait controlling under various ground conditions simulation environment. After that, generation capacity generators compared with each other. It shown inclusion layer provides smooth transition control between stance flight motion phases $$L_2$$ regularization beneficial performance. Finally, proposed network found be more successful controllers than generator, generate used training.