Research
My research revolves around building solutions for complex single- and multi-robot planning problems, and ranges from low-level path planning to high-level coordination strategies. My works encompass the use of a variety of different techniques drawn from operations research and combinatorial optimization (hardness reductions, approximation algorithms, mathematical programming), “classical” AI (search algorithms, symbolic planning), and machine learning (deep neural networks, Gaussian processes, reinforcement learning). Below, you can find some projects I've worked on during my PhD at the Politecnico di Milano and postdocs at Cornell and MIT. At Dexterity, I am working on decision-making algorithms for random, mixed-case robotic pallet building and truck loading. You can find some cool videos on Dexterity's website and on this Wired article.
Hierarchical Planning for Heterogeneous Multirobot Teams
During this project, I developed a hierarchical planner for a complex multi-robot task allocation problem in which tasks correspond to routing problems defined on separate areas of a given environment. The planner uses a Graph Neural Network to estimate the performance of a specific subteam on a given area. This estimate is used by the planner to inform the search.
Robotic Path Planning under Map Uncertainty
In this project, I developed a path planner able to drive a mobile robot in challenging outdoor scenarios where a full map of the environment might not be available in advance, which also includes uncertainty about the terrain semantics. The planner leverages a combination of classical planning techniques (RRTs, next-best-view planning) and machine learning (deep Bayesian neural networks) to plan multiple path hypotheses and iteratively reduce the corresponding uncertainties about the terrain semantics.
Building Communication Maps with Multirobot Teams
This project is about using a multi-robot system to build a communication map of an environment. Using Gaussian Processes as a regression tool, robots are able to predict, with a confidence value, the received signal strength between any two locations. This is a promising direction to try to overcome the usage of over-conservative (e.g. line-of-sight) or too simplistic (e.g. disk) communication models.
Computational Complexity of Multirobot Path Planning Problems
This project is about studying the computational complexity of graph-based multi-robot path planning problems. My most important results are the NP-completeness proofs of time-optimal multi-robot path planning on 2D grids and the PSPACE-completeness proof of connected multirobot path planning on graphs. The latter holds even when robots are unlabeled (i.e., any robot can reach any goal, provided that all goals are reached) and the physical environment is a 2D grid.
These projects are/were carried out in collaboration with some amazing researchers: Francesco Amigoni, Kavita Bala, Nicola Basilico, Mark Campbell, Yutao Han, Seth Hutchinson, Hubert Lin, Andrew Messing, Alberto Quattrini Li, Ioannis Rekleitis, Alessandro Riva, Nicholas Roy, Davide Tateo. Other current collaborators are: Beatriz Arruda Asfora, Martina Stadler, and Brian Wang.