Risk-aware Model Learning for Robot Control

This thesis addresses the challenge of improving robot tracking accuracy when the available mathematical models are imperfect. We propose a framework that combines physics-based models with data-driven learning to bridge this gap. First, we develop a fast tracking controller that uses learned model updates while respecting robot torque limits. Second, we introduce a structured learning method that captures the physical properties of robot motion, ensuring more reliable predictions even with limited data. Finally, we propose a risk-aware strategy to collect informative data without exposing the robot to unpredictable movements. Together, these contributions enable robots to learn and improve safely.