Implementation of an adaptive, model free, learning controller on the Atlas robot

Abstract

Recent events in natural and man-made disasters have highlighted the limitation in man's ability to confine and mitigate damage in such scenarios. Therefore, there is an urgent need for robotic technology that can function in all environments and serve as a substitute to humans in disaster scenarios. This paper presents research efforts to advance walking technology of humanoid robots with application to the Boston Dynamics Atlas robot. The Atlas was designed as part of the DARPA Robotics Challenge (DRC). The paper contribution is in a model free, walking trajectory tracking controller that is tested using GAZEBO robotics simulator. Artificial neural networks are used to learn the robot's nonlinear dynamics on the fly using a neuroadaptive control algorithm. The learned nonlinear dynamics are utilized along with a filtered error signal to generate input torques to control the system. Results show that the ability to approximate the robot nonlinear dynamics allows for full-body control without the need of modeling such a complex system. This ability is what makes the control scheme utilized appealing for complex, real-life, robotic applications that occur in a non-laboratory setting.