Obstacle Avoidance for Robot Arm by Conditional Generative Adversarial Networks

Abstract

Robots working at warehouses or living environments must avoid obstacles including people and robots itself. A lot of conventional researches on obstacle avoidance have problems that paths must be designed by the engineer or that the computational cost of path planning is high. In this research, we propose a novel obstacle avoidance without any prior design or the high computational cost. Based on manifold hypothesis, we implemented and trained Generative Adversarial Networks (GAN) to obtain lower dimensional representations from postures of 2 link arm robot which avoid obstacles. We confirmed the adjacency of the latent representation corresponds to that of joint angle after training of GAN. We also confirmed that when the trajectory on latent representations was drawn, it was mapped on the joint angle space and the smooth trajectory avoiding obstacles was planned.