Underwater robotics is revolutionizing our understanding of the oceans, surpassing traditional methods in investigating ocean phenomena. A recent study published in Science Robotics showcases the groundbreaking research in this field, highlighting the effectiveness of underwater robots in providing valuable in-situ data that complements satellite observations.
These robotic devices have the remarkable ability to reach extreme ocean depths of up to 4,000 meters, enabling scientists to explore intricate phenomena such as the CO2 absorption of marine organisms – a critical aspect of climate change mitigation. By employing reinforcement learning, a well-established concept in control and robotics, as well as natural language processing tools like ChatGPT, researchers have taught underwater robots how to optimize their actions in real-time to achieve specific objectives. Astonishingly, the success of these AI-driven action policies has even exceeded that of traditional methods based on analytical development in certain situations.
Ivan Masmitja, the lead author of the study, who conducted research at ICM-CSIC and MBARI, commented on the significance of this learning methodology, stating, This type of learning allows us to train a neural network to optimize a specific task, which would be very difficult to achieve otherwise. We have demonstrated that it’s possible to optimize the trajectory of a vehicle to locate and track objects moving underwater.
This pioneering research opens the door to a more comprehensive understanding of ecological phenomena in marine environments, such as the migration or movement patterns of various marine species. Joan Navarro, a researcher at ICM-CSIC involved in the study, underlined the potential for real-time monitoring of other oceanographic instruments through a network of autonomous robots.
The research team employed range acoustic techniques to estimate the location of underwater objects, utilizing distance measurements gathered from multiple points. By incorporating AI, particularly reinforcement learning, the researchers were able to identify the best data collection points, enabling the robot to follow the most optimal trajectory.
The Barcelona Supercomputing Center (BSC-CNS), home to Spain’s most powerful supercomputer, played a crucial role in training the neural networks. Professor Mario Martin from the UPC’s Computer Science Department emphasized the significant contribution of the supercomputer in expediting the parameter adjustment process of various algorithms.
After completing the training phase, the algorithms were rigorously tested on various autonomous vehicles during experimental missions in Sant Feliu de Guixols and Monterey Bay. One of the vehicles used was the AUV Sparus II, developed by VICOROB. The testing phase was facilitated by the Bioinspiration Lab at MBARI, headed by Kakani Katija.
Narcis Palomeras from UdG highlighted that the simulation environment simulated the control architecture of real vehicles, allowing for efficient implementation of the algorithms before deploying them at sea.
Looking ahead, the research team plans to explore the applicability of these algorithms in solving more complex missions, which includes deploying multiple vehicles to detect temperature changes in water (thermoclines), locating objects, or detecting algae upwelling using multi-platform reinforcement learning techniques.
This study was made possible thanks to the European Marie Curie Individual Fellowship awarded to researcher Ivan Masmitja in 2020, along with the BITER project, supported by the Ministry of Science and Innovation of the Government of Spain. The work of these dedicated researchers contributes significantly to our understanding of the oceans and the development of advanced technologies that can further our knowledge of marine ecosystems.
