23 out 2020
15:00 Master's Defense Fully distance
Theme
Deep Reinforcement Learning for BipedalLocomotion
Student
Yuri Corrêa Pinto Soares
Advisor / Teacher
Esther Luna Colombini
Brief summary
Robotics and their service applications with bipedal robots have recently expanded due to the possibility of using robots of this category in environments originally planned for human operation. However, bipedal locomotion has been shown to be challenging and practical due to the high dimensionality of the problem, since the action of andartypically involves the precise real-time control of multiple actuators and sensors in conjunction with complex dynamic systems. At the same time, reinforcement learning (RL) and its version with deep neural networks (DRL) are becoming a prominent approach to solving such problems, due to their ability to handle continuous, model-free processes. In this work, we model the locomotion problem as reinforcement learning, proposing a practical representation based on MPDs and multiple generic strategies for reinforcement functions. Then, we continued to develop a framework to integrate our choice simulator (CoppeliaSim) with the standard current interface for reinforcement learning (OpenAI Gym). Finally, we applied state-of-the-art algorithms in deep reinforcement learning with our framework in configurable experiments to validate our modeling and learn a stable walking policy in simulation for Marta's robot, a sophisticated humanoid robot with 25 degrees of freedom.
Examination Board
Headlines:
Esther Luna Colombini IC / UNICAMP
Marcos Ricardo Omena de Albuquerque Maximo ITA
Eric Rohmer FEEC / UNICAMP
Substitutes:
Leandro Aparecido Villas IC / UNICAMP
Alexandre da Silva Simões ICTS / UNESP