End-to-End Visual Obstacle Avoidance for a Robotic Manipulator using Deep Reinforcement Learning

• Ano de defesa: 2021
• Autor(a) principal: Sanches, Felipe Padula
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado por reforço profundo; Controle de movimento; Deep reinforcement learning; Desvio de obstáculos; Manipuladores robóticos; Motion control; Obstacle avoidance; Robot manipulators; Robot vision; Visão robótica
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-30082021-100712/

Recent changes in industrial paradigms enforce that robots must be intelligent and capable of decision-making. Robotic manipulators need to satisfy many requirements for operating properly. Perhaps the most fundamental one is the capability of operating in its environment without collisions. In this work, we perform visual obstacle avoidance on goal-reaching tasks of a robotic manipulator using an end-to-end Deep Reinforcement Learning model. The motion control policy is responsible for reaching a target position while at the same time avoiding an obstacle positioned randomly in the scene. This policy uses vision and proprioceptive sensor data to operate. We train the reinforcement learning agent using Twin-Delayed DDPG (TD3) algorithm in a simulated environment, utilizing the Unity game engine and the ML-Agents toolkit. Experiments demonstrate that the agent can successfully learn a meaningful policy to avoid obstacles using images.