Advances in embodied control: from Bayesian exploration to interactive full-body motion imitation

Thesis

Recent advancements in artificial intelligence, driven by machine learning, have revolutionized fields from computer vision to natural language processing. However, the physical capabilities of embodied agents, such as robots and virtual entities, still lag behind the agility of natural beings. This thesis explores the potential of learning-based methods and control theory to enhance the physical capabilities of artificial agents.

We develop Bayesian model-free RL approaches that quantify agent uncertainty, enabling more stable learning and improved exploration capabilities. We then explore the role of hierarchy and information asymmetry in model-free transfer learning, introducing a novel method for automating information asymmetry selection that enhances performance in simulated robot manipulation. Finally, we examine the use of keyframe motions as a source of priors and develop a model-predictive control method for interactive physics-based full-body motion imitation.

This thesis contributes novel algorithms, empirical analyses, and practical insights into the interplay between exploration, transfer learning, and motion imitation for embodied RL agents. These contributions collectively aim to advance the general field of RL and broaden the application spectrum of embodied agents in complex physical tasks.

Authors: Hartikainen, K

• Alternative title: Alternative title
• DOI: 10.5287/ora-jne2mvm05
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: reinforcement learning; embodied intelligence