Deep Reinforcement Learning for Dynamic Locomotion of Humanoid Robots

The research project is focused on using deep reinforcement learning to
solve dynamic locomotion of humanoid robots. In the past, locomotion is
mainly done using conventional analytical approaches, e.g.
Model-Predictive Control, which are limited because they require human
effort and knowledge, and demand high computing power to run online. In
contrast to on-line computation methods, the computation for machine
learning approaches can be outsourced offline. By doing so, faster
online performance for high dimensional control systems, such as
humanoids, can be achieved. Machine learning approaches such as deep
reinforcement learning also requires less human effort to design
compared to analytical approaches.

Given the increasingly more powerful deep RL algorithms, an increasing
number of research works have used deep RL to solve control tasks, as
the recent progress in deep RL algorithms designed for continuous action
domain has brought forward the possibility to apply reinforcement
learning continuous control tasks that involve complicated dynamics.

The project plans to explore the feasibilities of using deep
reinforcement learning to acquire bipedal control policies comparable or
better than analytical approaches while using less human effort. And
eventually design a control framework based on deep RL that is capable
of learning a wide range of balancing and walking strategies with
minimum human intervention, and compare the performance of the designed
control framework with current human labour intensive analytic
engineering approach. The project will also explore other methodologies
related to reinforcement learning, such as imitation learning and
transfer learning, and implement them into the proposed control
framework.