Open Collections

Ling, Hung Yu

University of British Columbia

Contemporary computer animation research has benefited substantially from the advancement of deep learning and deep reinforcement learning methods in the past decade. Despite the performance and flexibility of learning-based methods, significant manual effort is still required to tune the training data, algorithms, and environments, especially when computing resources are limited. In this thesis, we develop and evaluate a variety of learning-based methods that enable new skills, and that improve the stability of the learning algorithms and the quality of the synthesized motions. First, we present a framework for learning autoregressive kinematic motion generators and controllers from motion capture data. By disentangling the motion modelling and control tasks, our framework can efficiently synthesize controllable virtual characters by leveraging the strengths of supervised and reinforcement learning. Second, we study the effects of symmetry in learning physics-based locomotion controllers for bipedal characters. We evaluate four possible methods to impose symmetry and show that enforcing symmetry improves the naturalness and task performance of the trained controllers. Third, we explore the role of learning curricula in solving challenging physics-based stepping-stone tasks. The learning is significantly more robust and efficient under a learning curriculum which gradually increases the task difficulty. Finally, we combine simplified models and imitation learning to train brachiation controllers. We show that sparse task objective alone is sufficient for training a controller in an abstracted point-mass brachiation environment. Then, using the point-mass as a reference trajectory for the centre-of-mass, we can learn a control policy for a physics-based 14-link planar articulated gibbon model.

Text

2022-08-11

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/82325

Computer Science

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/