MACHINE LEARNING APPROACHES TO NAVIGATE TURBULENCE
Abstract
Navigating real environments poses the dramatic challenges of interpreting noisy and sparse signals coming from the physical world. Algorithms that would be able to solve this challenge are key to many real world applications required to function in uncontrolled environments: from search and rescue to demining and patrolling. While much is known on navigation in smooth environments, these approaches fail in unpredictable environments, for example dominated by turbulence. This project aims at using machine learning to elucidate the computations needed to extract useful information from turbulent stimuli and navigate to a target. A key observation is that sensory targets give out multiple cues regarding their whereabouts. We will leverage machine learning as a powerful tool to extract useful information from data. Massive dataset of sensory cues will be produced using state of the art numerical simulations of the propagation of chemical and mechanical signals from a target in the surrounding fluid. This dataset will be used to train machine learning algorithms first to elucidate what are the most informative stimuli to infer location of the target and what are the computations that are needed to extract this information. This information will feed reinforcement learning algorithms where agents will learn how to navigate from experience derived through repeated trials. The results provide an unprecedented connection between the physics of complex real-world stimuli; the mechanisms to extract information from these signals and the algorithms for navigating complex unpredictable environments.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 11, 2021
- Source ID
- FA86552017028
Entities
People
- Lorenzo Rosasco
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Genoa