Transferring human movements from videos to robots with Deep Reinforcement Learning

• Ano de defesa: 2022
• Autor(a) principal: Lessa, Nayari Marie
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Estadual Paulista (Unesp)
• Palavras-chave em Português: Inteligência Artificial; Imitação; Movimento; Robôs Dinâmica; Robots Dynamics; Imitation; Motion; Artificial intelligence
• Link de acesso: http://hdl.handle.net/11449/235583

The study of humanoid robots in the field of robotics has grown in recent decades in the direction of developing robots able to support humans in many applications. The evolution of machine learning techniques, particularly the Rein- forcement Learning (RL) approach, expanded the robotics domains to many new applications, based on the strategy to reinforce the agent through its interactions with the environment. Deep Reinforcement Learning (DRL) came to improve the RL technique allowing the application of robotics in highly complex task and scenarios. However, this ap- proach is well known for two major disadvantages: i) its high computational cost; ii) the difficulty in training the robot to achieving particular policies that are usually very difficult to model. Recently, RL approaches based on the imitation of reference movements have emerged in the robotics scenario. The learning process in this approach is based on the strategy of observing a reference movement policy from an expert and transfer it to the real robot with the maximum possible fidelity using DRL. In order to investigate this complex scenario, this work proposes an imitation process with three phases: i) the poses estimation of a human expert based on a video of this human performing a particular tasks; ii) the generation of reference motion trajectories to a robot; iii) the robot’s training in a simulated environment based on DRL technique to adapt and improve the reference movements to the new body scheme and dynamics of the robot. The investigation conducted with the Marta robot in a complex simulated environment showed that the imitation-based technique is able to make the robot kick a ball an average distance of 1m from YouTube videos.