Learning & Control
This pillar focuses on learning and control methods that allow robots to act robustly in the real world, especially when simulations, training conditions, and deployment environments do not perfectly match. I work on machine learning approaches that help robots adapt their behavior over time, transfer policies safely from simulation to reality, and maintain performance under changing conditions. The long-term objective is to support increasingly autonomous systems that can make reliable decisions in industrial and assistive contexts while remaining aligned with safety and operational constraints.
Key open problems include how to reduce the brittleness of learning-based control, how to maintain performance under continual change, and how to combine adaptation with safety guarantees. Current directions include continual domain adaptation, sim2real transfer, interaction-aware reinforcement learning, and methods that improve robustness without requiring excessive hyperparameter tuning or hand-crafted task-specific engineering. I am particularly interested in learning frameworks that are not only effective but also transparent enough to support engineering trust, iteration, and deployment in real collaborative systems.
