Abstract
This 5-hour module provides a comprehensive introduction to active learning, black-box optimization, and sample-based predictive control applied to modern robotics. The first half focuses on building surrogate models for expensive cost functions. Students will explore the mathematical foundations of Gaussian Processes as non-parametric function generators and learn how to apply Bayesian Optimization to balance exploration and exploitation through various utility functions (e.g., Expected Improvement, Maximum Probability of Improvement). The second half transitions to optimal control, detailing how Model Predictive Path Integral (MPPI) utilizes random sampling and GPU-parallelization to achieve fast planning and robust execution. Practical examples will demonstrate how these algorithms are successfully deployed on complex, agile platforms like quadrupeds and quadrotors.
The lecturer
Valerio Modugno has been a Lecturer in the Robotics and AI groups at UCL's Department of Computer Science since 2024. His research interests comprise Humanoid Whole-Body Control, Optimal Control, teleoperation for legged robots, Reinforcement Learning, Black-Box Optimization, and safety for control and learning strategies. Before joining UCL in 2022, he was a Post-Doctoral Researcher at Sapienza University of Rome, where he also received his Ph.D. in 2017. During his studies, he was a visiting researcher at the Technical University of Darmstadt in 2014 and at INRIA Grand-Est Nancy in 2015.
Where and when
© 2026 | Università degli studi di Verona
******** CSS e script comuni siti DOL - frase 9957 ********
