About

Roberto Furfaro is a Professor of Systems and Industrial Engineering at the University of Arizona. He serves as the Director of the Space Situational Awareness program and the Deputy Director of the S4 Space Center. He is a member of the Graduate Faculty in the Department of Aerospace & Mechanical Engineering. His professional roles involve research and leadership in space situational awareness, contributing to the university's focus on aerospace and mechanical engineering. His contact information includes a phone number and email address at the university.

Research topics

• Computer Science
• Artificial Intelligence
• Aerospace engineering
• Mathematics
• Engineering
• Physics
• Mathematical optimization
• Simulation
• Mathematical analysis
• Astrobiology
• Applied mathematics
• Algorithm
• Geography
• Geodesy
• Geology
• Astronomy
• Remote sensing
• Aeronautics

Selected publications

• Koopman-based Modeling for Rocket Landing Guidance Applications

  2026-01-08

  articleSenior author

  In the last years, the need for an efficient way to solve precisely the rocket landing guidance problem has emerged as crucial for reusability purposes. In this context, Koopman Operator Theory is promising because of its ability to globally linearize dynamical systems by lifting the state into a subspace of observables. More recent studies show that it is beneficial, in terms of accuracy, to learn a dictionary of observables through neural network optimization. In this work, we address the atmospheric, fuel-optimal, rocket landing problem to assess if a Koopman model can be used to efficiently solve the corresponding Optimal Control Problem. The main contributions of this work are the extension of dictionary learning techniques to include control and the formulation of a new framework to estimate a bilinear control model, after obtaining the equivalent linear modeling of the free dynamics. We show that dictionary learning methods provide more accurate linear control models than state-of-the-art Koopman-related methods, such that they can be used to efficiently build and solve a representative Linear Programming Problem. We also show that the adoption of a bilinear control model is possible, but this approach leads to a significant decrease in the related computational efficiency.

  PublisherDOI

• LLMs in the Loop: AI Agents for Precision Telescope Command

  2026-01-08

  articleSenior author

  Operating ground-based telescopes requires coordinating multiple software systems for orbital prediction, scheduling, execution, and data retrieval, tasks that are often labor-intensive and prone to error. We present an agentic framework that uses Large Language Models (LLMs) to translate natural-language observation requests into structured telescope commands. The agent interacts with a Telescope Automation Pipeline through Model Context Protocol (MCP) tools that provide access to telescope metadata, satellite-pass prediction, observation submission, and astrometric reporting. A web interface exposes the agent’s reasoning steps and tool calls, enabling transparent, conversational control of the full observation workflow. We evaluate several contemporary LLMs across representative telescope-operations scenarios. Models such as GPT-4o, Claude Opus 4.5, and Kimi K2 reliably executed required tool sequences, while others struggled with correct argument generation, preventing task completion. Results show that tool-use reliability, rather than general language ability, is the primary determinant of successful autonomous operation. This framework offers a path toward intuitive, low-overhead telescope control and future extensions to real-time analysis and adaptive scheduling.

  PublisherDOI

• Pontryagin Neural Networks for High-Fidelity Cislunar Orbital Transfers

  Journal of Guidance Control and Dynamics · 2026-03-08

  articleSenior author

  This paper leverages Pontryagin Neural Networks (PoNNs) to solve minimum-time low-energy low-thrust cislunar optimal transfers in a high-fidelity ephemeris model. PoNNs are neural networks tailored for solving two-point boundary value problems derived from the application of the Pontryagin Minimum Principle to optimal control problems (OCPs). Within PoNNs, the authors employ the Extreme Theory of Functional Connections to approximate state and costate using the constrained expressions to analytically satisfy the boundary conditions. The problems investigated regard a small spacecraft initially on a ballistic lunar transfer (BLT) directed to an elliptic lunar polar orbit, and a high elliptic earth orbit, transferred then to a series of BLTs reaching the Lunar Gateway. To solve these OCPs, the authors blend the PoNN classical framework with the particle swarm optimization, the penalty method, and some algebraic manipulations to analytically solve the transversality conditions arising from the variability of the departure orbit. As a result, the authors demonstrate that the PoNNs employed can effectively learn the underlying dynamics of the problem by approximating the positions and velocities with errors of [Formula: see text] and lower with respect to the scale of the model.

  PublisherDOI

• Meta-Reinforcement Learning for Conflict-Aware Sensor Scheduling in Space Surveillance and Tracking

  Virtual Community of Pathological Anatomy (University of Castilla La Mancha) · 2026-01-01

  article
  Publisher

• AI for space: theories, models and applications

  Neural Computing and Applications · 2025-07-11

  articleOpen access
  PublisherOA PDFDOI

• Action Chunking with Transformers for Image-Based Spacecraft Guidance and Control

  ArXiv.org · 2025-09-04

  preprintOpen access

  We present an imitation learning approach for spacecraft guidance, navigation, and control(GNC) that achieves high performance from limited data. Using only 100 expert demonstrations, equivalent to 6,300 environment interactions, our method, which implements Action Chunking with Transformers (ACT), learns a control policy that maps visual and state observations to thrust and torque commands. ACT generates smoother, more consistent trajectories than a meta-reinforcement learning (meta-RL) baseline trained with 40 million interactions. We evaluate ACT on a rendezvous task: in-orbit docking with the International Space Station (ISS). We show that our approach achieves greater accuracy, smoother control, and greater sample efficiency.

  PublisherOA PDFDOI

• Physics-informed orbit determination for X-GEO space situational awareness

  Acta Astronautica · 2025-09-12 · 2 citations

  articleSenior authorCorresponding
  PublisherDOI

• Meta-reinforcement learning guidance, navigation, and control for autonomous lunar landing with safe site selection

  Neural Computing and Applications · 2025-05-15 · 1 citations

  articleSenior author
  PublisherDOI

• A Pontryagin Neural Network application to tracklets correlation of optical observations

  Acta Astronautica · 2025-09-22

  articleOpen access

  As activity occurring in the region beyond the geostationary belt continues to grow, the number of satellites and space debris in this area is expected to rise significantly. As a direct consequence, the need of reliable methods to identify, correlate and catalog objects in this region will play a pivotal role to ensure the safety of space operations. This paper proposes a method based on Pontryagin Neural Networks (PoNNs), a specialized type of Physics-Informed Neural Networks (PINNs) designed to solve optimal control problems. By applying the Extreme Theory of Functional Connections, which combines the advantages of PINNs with the Theory of Functional Connections, the PoNN framework approximates the problem’s unknowns through a single-layer feed-forward neural network. The correlation problem is addressed by formulating an energy-optimal control problem that links two tracklets, under the assumption that the objects are non-maneuvering. When a successful correlation is achieved, the resulting solution corresponds to a ballistic trajectory that minimizes control effort. The Mahalanobis distance is then used to evaluate the correlation, which includes residuals on the data, the computed fuel cost associated to the optimal trajectory and the residuals on the physics. As a secondary goal, Initial Orbit Determination capabilities of the method are also investigated. The developed algorithm is validated using both simulated and real angles-only observations of objects governed by a two-body dynamical model. The real optical measurements, consisting of right ascension and declination data, were supplied by the telescopes operated by the Space4 Center at The University of Arizona. • Pontryagin Neural Networks applied to optical tracklets correlation. • No initial guess is required for the PoNN framework. • Validation of the PoNN framework with real observations provided by the Space4 Center. • Method suitable for Initial Orbit Determination purposes.

  PublisherDOI

• Optical Bench Validation of Meta-Reinforcement Learning for Image-Based Spacecraft Guidance and Navigation

  2025-03-21

  preprintSenior author

  This paper focuses on testing and validating meta-reinforcement learning algorithms for image-based spacecraft guidance, navigation, and control (GNC) using an optical bench designed to mimic small spacecraft’s computing hardware and camera systems. The optical bench setup is described in detail, including the workflow for GNC network deployment and the image processing techniques used to correct and enhance raw camera images. The analysis assesses, across one toy problem and two real-world space mission scenarios, the effectiveness and robustness of the GNC network when processing actual camera images in place of the simulated images used for the network training. Also, it aims to evaluate the network computational performance on typical onboard computing hardware.

  PublisherDOI

Frequent coauthors

• B. D. Ganapol

  45 shared

• Brian Gaudet

  43 shared

• Paolo Picca

  41 shared

• Enrico Schiassi

  CNH Industrial (Italy)

  40 shared

• Andrea Scorsoglio

  University of Arizona

  37 shared

• V. Reddy

  35 shared

• Andrea D’Ambrosio

  University of Arizona

  34 shared

• Mario De Florio

  29 shared