Footstep planning of humanoid robot in ROS environment using Generative Adversarial Networks (GANs) deep learning

Abstract

This paper proposes deep learning-based footstep planning using Generative Adversarial Networks (GANs) for the indoor navigation of humanoid robots. The GAN-based architecture presents an efficient and accurate path to plan the footsteps of a humanoid robot on Robot Operating System (ROS) based architecture. During navigation, humanoid robots must understand their surroundings and be able to generate footsteps accurately. Although some algorithms that are based on sampling, such as Rapidly Exploring Random Tree (RRT*) and A*, are widely used for path planning, they fail to perform in narrow paths, especially for the footstep generation of humanoid robots. The widely growing deep learning approaches such as GANs are now producing extremely surprising results in solving real-life problems. The experiments conclude that GAN based approach performs better than conventional Dijkstra’s or A* algorithms. The accuracy of the generated footsteps from the GAN-based path planner comes out to be approximately 93%.