Safe and global optimality robot control - Beyond control invariance and reinforcement learning

ABSTRACT
In recent years, advanced data-driven control methods are unlocking the potential of complex robotics systems, and we can expect this trend to continue at an exponential rate in the near future. However, these methods present two major shortcomings. The resulting policies cannot ensure constraint satisfaction (i.e. safety), and their computation relies on brute-force sampling-based methods. In this talk I will discuss how we are using tools from optimal control to address these issues. 

To ensure safety, a well-known tool is the control-invariant set (a.k.a. safe set). Unfortunately, for nonlinear systems, such sets can only be approximated, and classic methods break down if the set is not exact. I will present some novel control schemes that guarantee safety under weaker assumptions than classic methods. We evaluated our approaches on simulated and real robot manipulators, empirically demonstrating that they lead to less constraint violations, while retaining good performance (i.e. tracking cost) and low computation times.

To improve the efficiency of sampling-based data driven methods, I will show you how we have embedded gradient-based Trajectory Optimization inside an Actor-Critic scheme. Our novel GPU-based implementation is up to 10 times faster than PPO in our benchmarks, and it has been validated through experiments on a real quadruped robot.

BIO

Andrea Del Prete is an Associate Professor in the Department of Industrial Engineering at the University of Trento (Italy), where he has been serving since 2022. He is a member of the Interdepartmental Robotics Lab (IDRA). From 2019 to 2021, he was a tenure-track Assistant Professor (RTD-B) in the same department.
In 2018, he was a Research Scientist in the Movement Generation and Control Group at the Max Planck Institute for Intelligent Systems (Tübingen, Germany), led by Ludovic Righetti.
From 2014 to 2017, he was an Associate Researcher in the Gepetto Team at LAAS-CNRS (Toulouse, France), where he worked extensively on the humanoid robot HRP-2. During this period, he collaborated closely with Nicolas Mansard, Olivier Stasse, Steve Tonneau, and Justin Carpentier.
Prior to joining LAAS, he spent four years at the Italian Institute of Technology (IIT, Genoa, Italy)—three as a Ph.D. student and one as a postdoctoral researcher—working with Lorenzo Natale and Francesco Nori on the iCub humanoid robot.
Andrea Del Prete has been serving as an Associate Editor for IEEE Robotics and Automation Letters since 2020 and for IEEE Transactions on Robotics since 2024.