Aprendizado por reforço profundo para locomoção bípede

Author(s)

Yuri Corrêa Pinto Soares

• Document type: Master's Dissertation
• Press: Campinas, SP.
• Institution: Universidade Estadual de Campinas (UNICAMP). Instituto de Computação
• Defense date: 2020-10-23

Examining board members

Esther Luna Colombini; Marcos Ricardo Omena de Albuquerque Maximo; Eric Rohmer

• Advisor: Esther Luna Colombini

Abstract

Robotics and its service applications with biped robots have faced an upsurge lately asthis category of robots is suitable for deployment in environments originally designed foroperation by humans. However, bipedal locomotion has proven to be a challenge in theoryand practice due to the problem¿s high dimensionality: as walking gaits typically involveprecise real-time control of multiple actuators and sensors, coupled with complex dynam-ical systems. Concomitantly, reinforcement learning (RL) and its deep neural networkversion (DRL) are becoming a prominent approach in solving such challenging controlproblems due to their capacity to work on continuous and model-free processes. In thiswork, we modeled a locomotion task as an RL problem by proposing practical MDP repre-sentations and generalizable reward engineering strategies. We then proceeded to developa framework for integrating our simulator of choice (CoppeliaSim [11]) with the de factostandard interface for Reinforcement Learning (OpenAI Gym [5]). Finally, we appliedstate-of-the-art DRL algorithms within our framework in configurable and reproducibleexperiments to validate our modeling and learn a stable walking gait in simulation for theMarta robot, a sophisticated humanoid robot with 25 DOFs (AU)

• FAPESP's process: 17/21426-7 - Deep reinforcement learning for bipedal locomotion
• Grantee: Yuri Corrêa Pinto Soares
• Support Opportunities: Scholarships in Brazil - Master