Abstract: 

Robots operating in the physical world must learn what to do from limited data while remaining reliable when reality differs from their models. This talk presents a probabilistic perspective that connects inverse optimal control, free-energy-based decision-making, and distributionally robust learning. First, we show how the latent, potentially non-convex and time-varying objectives underlying observed behavior can be reconstructed through convex optimization, even for nonlinear and stochastic systems. We then introduce a distributionally robust free-energy principle that explicitly accounts for uncertainty in the environment, producing policies that remain effective under discrepancies between training and deployment conditions. Finally, we combine this principle with maximum-diffusion learning to unify exploration and robustness in continuous-control robotics. The resulting framework learns both dynamics and rewards while providing explicit robustness guarantees against epistemic uncertainty. Simulations and hardware experiments demonstrate improved reliability and a reduced sim-to-real gap without task-specific fine-tuning. Together, these results suggest a path from interpreting observed behavior to learning and deploying dependable policies for robotic manipulation.

Biography:  

Hozefa Jesawada is a Postdoctoral Researcher at New York University Abu Dhabi. He earned his PhD in Information Technologies for Engineering from the University of Sannio, Italy, and subsequently worked as a postdoctoral researcher at the University of Salerno. Before his doctoral studies, he was a Machine Learning Researcher at the International Gemological Institute. He holds an M.Tech. in Control Systems from VJTI Mumbai and a bachelor’s degree in Electrical Engineering. His research lies at the intersection of control theory, machine learning, and robotics, focusing on learning-based and inverse optimal control, reinforcement learning, free-energy methods, and robustness guarantees for real-world robotic systems.