Intro
When PhD student Tiberiu-ioan Szatmari walks in the door at NASA’s Jet Propulsion Laboratory (JPL) for the first time on September 18th, 2023, it is a childhood dream come true. Tiberiu-ioan is a PhD student with Eriksholm Research Centre and he uses distributed optimization as part of his PhD work. With JPL also showing a keen interest in such methods, Tiberiu-ioan approached them to identify opportunities for a research stay.
About
“A project was defined over several meetings between researchers from JPL Robotics and me. It follows a ‘similar topics, different application’ approach, where the goal is to use federated learning techniques for multi-agent robotic exploration and mapping, related to the CADRE mission, scheduled for the Moon in late 2024. This way, it is related to both my project and interests, as well as JPL’s near-future goals.”
The project is named “Decentralized Multi-Agent Learning for Collaborative Mapping”, and the plan is to investigate federated learning techniques for the problem of joint exploration and multi-agent mapping. In particular, the project will use such techniques to learn traversability maps for a set of N-agents and investigate how federated learning can be used for reconciling and synchronizing the map being measured by each agent in an asynchronous fashion.
“Some of the techniques I’m investigating, such as federated learning, use models trained on local data to make a global “knowledge” pool, which is then shared back such that local models can use the new global pool to achieve better performance than only using local data. In my project this focuses on the privacy aspect for users, while for JPL it’s about improving the performance and adding new capabilities for a multi-rover mission scheduled to explore a new area of the Moon”, says Tiberiu-ioan.
Recently, multi-agent missions have received renewed interest at JPL as a key enabling technology for robot exploration. For example, the CADRE mission uses three surface rovers to map and carry out scientific measurements for Lunar exploration. A part of the ConOps for such missions includes the mapping phase during which hazards and regions of interest in the surrounding area are identified. For multi-agent missions, the maps collected by each agent must be synchronized to downlink a unified map representation and this is the primary area of investigation for this work.
“It is a childhood dream of mine to collaborate with NASA on space exploration-related projects, and I see this as an exciting addition to my PhD work”, ends Tiberiu-ioan.