Adrián Bañuls Arias

Short bio

Adrián Bañuls Arias received his Bachelor’s Degree in Robotics, Electronics and Mechatronics Engineering from the University of Málaga in 2019, graduating with honors and publishing a paper at the IEEE International Symposium on Safety, Security and Rescue Robotics (SSRR). He later earned two master’s degrees, one in Mechatronic Engineering and another in Artificial Intelligence. After completing his postgraduate studies, he spent several years in industry developing and integrating advanced robotic systems across various sectors, combining practical engineering expertise with cutting-edge AI methods. In 2025 he began his Ph.D. in Mechatronic Engineering at the University of Málaga, where his research focuses on Reinforcement Learning and its applications to robotics.

Research

Research interests

• Reinforcement Learning for robotics
• Control of Robotic Manipulators
• Mobile robot navigation

PhD Thesis

Title: Action Time in Reinforcement Learning for Robotics
Programme: PhD Programme in Mechatronic Engineering, University of Málaga
Supervisors: Juan Antonio Fernández Madrigal and Vicente Arévalo Espejo
Start: February 2025

Abstract. Reinforcement Learning (RL) has shown great potential in robotics over the past decade. However, one fundamental aspect remains largely overlooked: time. RL has traditionally been applied to turn-based games such as chess, where the problem reduces to a discrete sequence of moves, and time plays no role. In robotics, by contrast, time emerges in many ways, including the optimal duration of each action, the range of durations allowed by the operating frequencies of actuators and sensors, and the gap between nominal and effective action times caused by latencies, communication delays or computational load. This thesis studies how these temporal factors affect the performance of RL algorithms and how to adapt these algorithms accordingly.

The main contributions to date are as follows. First, a software library has been developed to decouple the agent from the learning algorithm, enabling a more reliable study of time. Second, an exhaustive multi-metric analysis has been conducted to determine the optimal action time for a robotic navigation task. Third, the robot has been given the ability to learn and select its own action time, which has led to a tool for building safety maps of its environment. Ongoing work extends this research to further temporal factors.

Publications