Researchers at the Institute of Research in Mechatronic Engineering and Cyberphysical Systems at the University of Malaga (IMECH.UMA) have introduced a novel approach to artificial intelligence that enables robots to learn not only what decision to make, but also when to make it, by incorporating time as a crucial factor in their learning process.
The project, known as TYRELL (Time in Reinforcement Learning for Robotics), is funded by the Ministry of Science, Innovation, and Universities, and addresses a commonly overlooked aspect in reinforcement learning, a widely used artificial intelligence technique to empower robots with autonomy. Unlike conventional systems where robots make decisions at fixed time intervals, the team at IMECH.UMA suggests adapting these intervals dynamically to suit each situation.
The principal investigator, Juan Antonio Fernández Madrigal, likens the robot’s decision-making process to how humans adjust their reactions while driving on a highway versus parking, highlighting the need for varying decision-making frequencies. The team argues that certain situations demand quick responses, while in others, maintaining a consistent action over a longer period can conserve energy and computational effort.
The researchers have devised ‘Subjective Time Maps,’ a new representation of the environment that assists robots in learning the appropriate decision frequency for each area. This innovation allows robots to respond promptly in complex situations and space out decisions in favorable environments, enhancing safety, navigation precision, and energy efficiency.
Apart from its scientific significance, the project promotes open science through the release of software libraries, simulators, and experimental datasets that encourage the replication and validation of findings by other research groups. Recent outcomes of this study published in the international journal ‘Computer Modeling in Engineering & Sciences’ showcase the optimization of autonomous navigation even in scenarios with uncertainties in robot localization or communication delays.
The TYRELL project brings together researchers from the University of Malaga, Örebro University (Sweden), and École Polytechnique Fédérale de Lausanne (EPFL) (Switzerland), exploring how a temporal dimension can enhance learning in mobile robots and robotic manipulators. The implications span diverse fields such as logistics, agriculture, industrial inspection, service robotics, and search and rescue missions.
In summary, this innovative research not only contributes to advancements in robotic learning but also holds promise for various practical applications across industries.
About the Author
