Date of Award
Doctor of Philosophy (PhD)
School of Computing
Imitating and predicting human motions have wide applications in both graphics and robotics, from developing realistic models of human movement and behavior in immersive virtual worlds and games to improving autonomous navigation for service agents deployed in the real world. Traditional approaches for motion imitation and prediction typically rely on pre-defined rules to model agent behaviors or use reinforcement learning with manually designed reward functions. Despite impressive results, such approaches cannot effectively capture the diversity of motor behaviors and the decision making capabilities of human beings. Furthermore, manually designing a model or reward function to explicitly describe human motion characteristics often involves laborious fine-tuning and repeated experiments, and may suffer from generalization issues. In this thesis, we explore data-driven approaches using generative models and reinforcement learning to study and simulate human motions. Specifically, we begin with motion synthesis and control of physically simulated agents imitating a wide range of human motor skills, and then focus on improving the local navigation decisions of autonomous agents in multi-agent interaction settings. For physics-based agent control, we introduce an imitation learning framework built upon generative adversarial networks and reinforcement learning that enables humanoid agents to learn motor skills from a few examples of human reference motion data. Our approach generates high-fidelity motions and robust controllers without needing to manually design and finetune a reward function, allowing at the same time interactive switching between different controllers based on user input. Based on this framework, we further propose a multi-objective learning scheme for composite and task-driven control of humanoid agents. Our multi-objective learning scheme balances the simultaneous learning of disparate motions from multiple reference sources and multiple goal-directed control objectives in an adaptive way, enabling the training of efficient composite motion controllers. Additionally, we present a general framework for fast and robust learning of motor control skills. Our framework exploits particle filtering to dynamically explore and discretize the high-dimensional action space involved in continuous control tasks, and provides a multi-modal policy as a substitute for the commonly used Gaussian policies. For navigation learning, we leverage human crowd data to train a human-inspired collision avoidance policy by combining knowledge distillation and reinforcement learning. Our approach enables autonomous agents to take human-like actions during goal-directed steering in fully decentralized, multi-agent environments. To inform better control in such environments, we propose SocialVAE, a variational autoencoder based architecture that uses timewise latent variables with socially-aware conditions and a backward posterior approximation to perform agent trajectory prediction. Our approach improves current state-of-the-art performance on trajectory prediction tasks in daily human interaction scenarios and more complex scenes involving interactions between NBA players. We further extend SocialVAE by exploiting semantic maps as context conditions to generate map-compliant trajectory prediction. Our approach processes context conditions and social conditions occurring during agent-agent interactions in an integrated manner through the use of a dual-attention mechanism. We demonstrate the real-time performance of our approach and its ability to provide high-fidelity, multi-modal predictions on various large-scale vehicle trajectory prediction tasks.
Xu, Pei, "Motion Synthesis and Control for Autonomous Agents using Generative Models and Reinforcement Learning" (2023). All Dissertations. 3407.
Author ORCID Identifier