Seminario Geometría, Mecánica y Control
Learning-Based Decentralized Navigation and Safety-Critical Control for Multi-Agent Robotic Systems
Ponente: Omayra Yago Nieto (Universidad Politécnica de Madrid)Fecha: viernes 16 de octubre de 2026 - 12:00Lugar: Aula Gris 3, ICMAT; Seminario de Geometría y Topología (ULL)
Online: Zoom. Meeting ID: 890 0727 4974; Access code: 159180
Resumen:
Safe coordination of multi-agent systems naturally combines geometric, probabilistic, and constrained dynamical problems: constructing decentralized vector fields with prescribed asymptotic behavior, accounting for uncertainty, and guaranteeing the forward invariance of collision-free regions. We develop a learning-based framework for decentralized navigation in three-dimensional environments. For fully actuated agents on SE(3), we construct decentralized navigation functions whose induced vector fields encode convergence and collision avoidance, while Gaussian Processes are used to model unknown dynamics and derive probabilistic tracking guarantees.
We then consider underactuated quadrotors, where these vector fields must be realized through thrust and attitude dynamics, leading to higher-order safety constraints formulated using robust high-order control barrier functions. Finally, we study the feasibility of these invariance conditions under bounded actuation, including loss of control effectiveness and incompatibility among multiple safety constraints, connecting geometric navigation, probabilistic learning, and constrained dynamical systems.
