gammelli at stanford dot edu·daniele dot gammelli at ai4i dot it
I am the Research Director of the Machine Intelligence for Robot Autonomy Laboratory at the Italian Institute of Artificial Intelligence (AI4I), and a Researcher in the Department of Aeronautics and Astronautics at Stanford University within the Autonomous Systems Lab. My lab develops the next generation of tools for Physical AI, advancing embodied intelligence and translating it into real-world impact across aerospace autonomy, future mobility systems, and autonomy at large.
I completed my PhD at the Technical University of Denmark (DTU), where I was fortunate to spend my time in the Machine Learning for Smart Mobility Lab.
Openings
Twitter / Google Scholar / Github / CV / LinkedIn
The goal of my research is to develop the algorithmic foundations and system-level methodologies that enable AI-powered autonomous systems to operate safely, efficiently, and reliably in high-stakes environments.
Our paper SAGES: Semantic Trajectory Generation for Goal-Oriented Spacecraft Rendezvous was selected as the Intelligent Systems Best Paper Award at the 2026 AIAA SciTech Forum!
The Principles of Robot Autonomy book is now complete and available online!
Our work on Multi-Timescale Model Predictive Control for Slow-Fast Systems was accepted at ECC 2026!
Our paper Graph Neural Model Predictive Control for High-Dimensional Systems was accepted at ICRA 2026! Introducing scalable and effective learning-based MPC for soft robotics.
Our paper Reproducibility in the control of autonomous mobility-on-demand systems was published in IEEE T-RO!
Another year of AA203: Optimal and Learning-based Control at Stanford begins!
Two IEEE RA-L papers accepted on learning-based control of multi-agent robotic systems: GRAND: Guidance, Rebalancing, and Assignment for Networked Dispatch in Multi-Agent Path Finding and Accelerating High-Capacity Ridepooling in Robo-Taxi Systems.
Very happy to host the 4th Workshop on Artificial Intelligence for Space at CVPR 2026! See the full call for papers.
Wonderful visit at NASA JPL! I gave a talk on “Advancing Aerospace Autonomy with Foundation Models”.
Our work on multi-modal transformers for spacecraft trajectory optimization was selected as single-track presentation at ACC 2025.
Our works on vision foundation models for anomaly detection and transformers for model predictive control were presented at ICRA 2025.
Our ICLR paper on Offline Hierarchical Reinforcement Learning via Inverse Optimization was presented in Singapore.
Another year of AA203: Optimal and Learning-based Control at Stanford begins!
Check out our two IEEE Aerospace Conference papers: Space-LLaVA: a Vision-Language Model Adapted to Extraterrestrial Applications and Towards Robust Spacecraft Trajectory Optimization via Transformers.
New pre-print: Robo-taxi fleet coordination at scale via reinforcement learning.
Autonomous systems are increasingly central to modern society, deeply embedded in critical infrastructure, human-facing services, and scientific progress. From resilient power grids and intelligent transportation networks to cyber-physical workforces that augment human capabilities and coordinated fleets of aerial, terrestrial, and orbital robots, autonomy is rapidly moving from research prototypes to indispensable real-world operations. While the exact embodiments, safeguards, and degrees of agency will evolve, one fact is certain: persistent, trustworthy, and general-purpose autonomy will be essential.
Recent advances in Artificial Intelligence (AI) have greatly accelerated this transformation, with machine learning (ML)-powered autonomy stacks now performing tasks previously considered beyond the capabilities of traditional, non-learning-based systems. Most notably, the emergence of internet-scale, broadly capable Foundation Models (FMs) offers an opportunity to fundamentally rethink how autonomous systems are designed, deployed, and operated. Trained on vast and diverse datasets, these models capture broad priors about the world and have achieved breakthroughs in vision, language, and multi-modal reasoning.
However, delivering reliable, general-purpose autonomy presents unique challenges for AI systems. Autonomous systems must guarantee safety and reliability in real-world environments, often under conditions that cannot be anticipated at design time. They must also operate within the constraints of on-board compute and learn from only scarce embodiment-specific data. These stringent requirements stand in sharp contrast to the properties of current AI systems, which—despite their remarkable capabilities—depend on vast amounts of training data and large-scale compute, and remain vulnerable to hallucinations and brittle generalization.
To that end, I address these challenges by employing and advancing techniques from AI/ML, control theory, and mathematical optimization. I apply the results to aerospace robotics, future mobility systems, and autonomy at large.
For an updated list of my publications, visit my Google Scholar page.
IEEE Aerospace Conference, 2025
We present Space-LLaVA, initialized from a pre-trained LLaVA 13B model and fine-tuned to extraterrestrial applications on tasks such as, e.g., instruction-following, conversations constructed from three extraterrestrial datasets.
International Conference on Learning Representations (ICLR), 2025
We propose OHIO: a framework for offline reinforcement learning (RL) of hierarchical policies. We demonstrate our framework on robotic and network optimization problems and show that it substantially outperforms end-to-end RL methods and improves robustness. We investigate a variety of instantiations of our framework, both in direct deployment of policies trained offline and when online fine-tuning is performed.
American Control Conference, 2025 Single-track Presentation
We present a novel trajectory generation framework that generalizes across diverse problem configurations, by leveraging high-capacity transformer neural networks capable of learning from multimodal data sources. Specifically, our approach integrates transformer-based neural network models into the trajectory optimization process, encoding both scene-level information (e.g., obstacle locations, initial and goal states) and trajectory-level constraints (e.g., time bounds, fuel consumption targets) via multimodal representations.
IEEE Aerospace Conference, 2025
This work extends the capabilities of ART to address robust chance-constrained optimal control problems. Specifically, ART is applied to challenging rendezvous scenarios in Low Earth Orbit (LEO), ensuring fault-tolerant behavior under uncertainty. The proposed warm-starting strategy is shown to consistently produce high-quality reference trajectories, achieving up to 30% cost improvement and 50% reduction in infeasible cases compared to conventional methods, demonstrating robust performance across multiple state representations.
IEEE Robotics and Automation Letters, 2025
We present a unified framework to combine the main strengths of optimization-based and learning-based methods for MPC. Our experiments, performed in simulation and the real world onboard a free-flyer platform, demonstrate the capabilities of our framework to improve MPC convergence and runtime. Compared to purely optimization-based approaches, results show that our approach can improve trajectory generation performance by up to 70%, reduce the number of solver iterations by up to 45%, and improve overall MPC runtime by 7x without loss in performance.
IEEE Aerospace Conference, 2024
We introduce the Autonomous Rendezvous Transformer (ART) for spacecraft trajectory optimization. ART combines optimization-based and AI-based methods, which improves task performance while providing the safety assurances needed for space operations. The method entails embedding high-capacity (namely, transformer-based) neural network models within the optimization process for trajectory generation.
International Conference on Machine Learning (ICML), 2023
We propose a learning-based framework to handle a broad class of network problems by exploiting the main strengths of graph representation learning, reinforcement learning, and classical operations research tools.
arXiv, 2023
We present the E-AMoD control problem through the lens of reinforcement learning and propose a graph network-based framework to achieve drastically improved scalability and performance over heuristics.
Conference on Knowledge Discovery and Data Mining (KDD), 2022 Oral
We formalize the multi-city AMoD problem through the lens of meta-reinforcement learning and devise an RL agent based on recurrent graph neural networks. In our approach, AMoD controllers are explicitly trained such that a small amount of experience within a new city will produce good system performance.
Transportation Research Part C: Emerging Technologies (TR-C), 2022
We devise a deep generative model to forecast future pickup and return rates for shared mobility services. We show how more accurate predictions do not necessarily translate into better inventory decisions. By providing insights into the interplay between forecasts, model assumptions, and decisions, we point out that forecasts and decision models should be carefully evaluated and harmonized to optimally control shared mobility systems.
Pattern Recognition, 2022
We propose a novel extension to the multi-output Gaussian process framework that leverages information from multiple correlated outputs to address the censoring problem. We further position the proposed model into a general framework capable of dealing with arbitrary likelihood functions for the purpose of censored modelling.
Pattern Recognition, 2022
ICML Workshop on Invertible Neural Nets and Normalizing
Flows, 2022
We introduce recurrent flow networks (RFN) for spatio-temporal data prediction by explicitly disentangling between temporal and spatial variability.
Conference on Decision and Control (CDC), 2021
ICML Workshop on RL for Real Life,
2021 Best Paper Candidate
We propose a deep reinforcement learning framework to control the rebalancing of AMoD systems through graph neural networks. Crucially, we demonstrate that graph neural networks enable reinforcement learning agents to recover behavior policies that are significantly more transferable, generalizable, and scalable than policies learned through other approaches.
Transportation Research Part C: Emerging Technologies, 2020
Transportation Research
Board Annual Meeting (TRB), 2020 Lectern Session
We propose a general method for censorship-aware modeling, for which we devise a censored likelihood function. We apply this method to the task of shared mobility demand prediction by incorporating the censored likelihood within a Gaussian Process model, which can flexibly approximate arbitrary functional forms.
The complete book — chapters, lecture recordings, and slides — is now
available online.
porabook.com →
