Diego Martínez Baselga defenderá su tesis doctoral titulada “Autonomous Navigation in Dynamic Environments using Geometry, Learning and Control” el próximo viernes, 23 de enero de 2026, a las 10:00 horas en el Salón de actos del Edificio Ada Byron.
La investigación ha sido dirigida por los profesores Luis Riazuelo y Luis Montano. El aborda los desafíos de la navegación autónoma en entornos dinámicos. Su investigación, bajo la dirección de Luis Riazuelo y Luis Montano, busca mejorar la capacidad de los robots para moverse de forma segura y eficiente en presencia de agentes en constante movimiento. La defensa tendrá lugar el viernes, 23 de enero de 2026, a las 10:00 horas, en el Salón de actos del Edificio Ada Byron, y también se podrá seguir a través de videollamada.
Resumen:
Autonomous robots are becoming increasingly important in our daily lives. They are be coming indispensable in warehouses, factories, and open spaces for tasks such as delivery, service, and support. Autonomous navigation is the ability to move from one location to another without human control. It is a fundamental requirement for robot autonomy. However, achieving safe, robust, and efficient motion planning remains a challenge in environments populated by dynamic agents such as humans, vehicles, and other robots. These agents introduce unpredictability, requiring navigation systems to be reactive and account for environmental dynamism. This thesis aims to advance the state of the art in autonomous navigation within dynamic environments. The first part focuses on Deep Reinforcement Learning (DRL) planners, which train deep neural networks to map sensory information to velocity commands through interaction with the environment. The work improves robustness, adaptability, and scalability through enhancements to motion planning components. The second part introduces a motion planner that integrates geometric collision avoidance with adaptive control to enable efficient and smooth navigation for multi-robot systems in dynamic settings. The planner adapts to the cooperation level of surrounding agents. In addition, we propose methods to let the planner work in presence of kinematic constraints and perception uncertainty. The third part introduces a motion planner that uses supervised learning to imitate the high-level navigation decision making process of humans (pass before/after, avoid left/right, etc.). This decisions are learned from real-world datasets of pedestrians walking in crowds, and are applied as guidance trajectories that a Model Predictive Control Planner (MPC) tracks. It also explores the use of Large Language Models and Visual Language Models to let users give queries to the robot (”go faster”, ”keep a safe distance to pedestrians”). Overall, this thesis contributes to improve autonomous navigation in dynamic environments to reduce the gap of robots and real-world deployment. The proposed methods have been extensively evaluated in simulated and hardware experiments, proving their effectiveness and efficiency to be applied in real platforms.
Enlaces:
https://sites.google.com/view/dmartnezbaselga
