Motor control and strategy discovery for physically simulated characters

In physics-based character animation, motions are realized through control of simulated characters along with their interactions with the virtual environment. In this thesis, we study the problem of character control on two levels: joint-level motor control which transforms control signals to joint torques, and high-level motion control which outputs joint-level control signals given the current state of the character and the environment and the task objective. We propose a Modified Articulated-Body Algorithm (MABA) which achieves stable proportional-derivative (PD) low-level motor control with superior theoretical time complexity, practical efficiency and stability than prior implementations. We further propose a high-level motion control framework based on deep reinforcement learning (DRL) which enables the discovery of appropriate motion strategies without human demonstrations to complete a task objective. To facilitate the learning of realistic human motions, we propose a Pose Variational Autoencoder (P-VAE) to constrain the DRL actions to a subspace of natural poses. Our learning framework can be further combined with a sample-efficient Bayesian Diversity Search (BDS) algorithm and novel policy seeking to discover diverse strategies for tasks with multiple modes, such as various athletic jumping tasks.