Exploring Robot Trajectories in Panoramic Vision-Based Control Using Deep Reinforcement Learning

Takieddine Soualhi, Nathan Crombez, Yassine Ruichek, Alexandre Lombard, Stéphane Galland

Abstract: In this paper, we study the problem of direct trajectories in visual control of nonholonomic mobile robots. To address this challenge, we propose combining deep reinforcement learning with panoramic vision to learn a control policy that maps input images to control velocities. We demonstrate that the adopted approach can precisely drive the robot to a desired pose. Additionally, it enables the emergence of various control strategies to achieve interesting trajectories. Our approach is validated and evaluated in simulation and transferred to the real world.