Learning dynamic motor skills for terrestrial locomotion
Item statusRestricted Access
Embargo end date30/11/2021
The use of Deep Reinforcement Learning (DRL) has received significantly increased attention from researchers within the robotics field following the success of AlphaGo, which demonstrated the superhuman capabilities of deep reinforcement algorithms in terms of solving complex tasks by beating professional GO players. Since then, an increasing number of researchers have investigated the potential of using DRL to solve complex high-dimensional robotic tasks, such as legged locomotion, arm manipulation, and grasping, which are difficult tasks to solve using conventional optimization approaches. Understanding and recreating various modes of terrestrial locomotion has been of long-standing interest to roboticists. A large variety of applications, such as rescue missions, disaster responses and science expeditions, strongly demand mobility and versatility in legged locomotion to enable task completion. In order to create useful physical robots, it is necessary to design controllers to synthesize the complex locomotion behaviours observed in humans and other animals. In the past, legged locomotion was mainly achieved via analytical engineering approaches. However, conventional analytical approaches have their limitations, as they require relatively large amounts of human effort and knowledge. Machine learning approaches, such as DRL, require less human effort compared to analytical approaches. The project conducted for this thesis explores the feasibility of using DRL to acquire control policies comparable to, or better than, those acquired through analytical approaches while requiring less human effort. In this doctoral thesis, we developed a Multi-Expert Learning Architecture (MELA) that uses DRL to learn multi-skill control policies capable of synthesizing a diverse set of dynamic locomotion behaviours for legged robots. We first proposed a novel DRL framework for the locomotion of humanoid robots. The proposed learning framework is capable of acquiring robust and dynamic motor skills for humanoids, including balancing, walking, standing-up fall recovery. We subsequently improved upon the learning framework and design a novel multi-expert learning architecture that is capable of fusing multiple motor skills together in a seamless fashion and ultimately deploy this framework on a real quadrupedal robot. The successful deployment of learned control policies on a real quadrupedal robot demonstrates the feasibility of using an Artificial Intelligence (AI) based approach for real robot motion control.