Publications
Peer-reviewed journals, conferences, and workshops.
2026
-
Sampling-Based Coordination-Informed Multi-Objective Multi-Robot Reinforcement Learning
IEEE Transactions on Robotics journal
Multi-robot systems must simultaneously optimize competing objectives while maintaining coordinated behavior. Existing multi-agent reinforcement learning approaches often rely on fixed or centralized coordination, which limits adaptability and violates distributed constraints. This work introduces the Coordination-Informed Multi-Objective Reinforcement Learning (CIMORL) framework, integrating a distributed weight prediction mechanism, a privileged expert training strategy, and theoretical guarantees for Pareto-optimal solutions. We present the base CIMORL method alongside two sampling-based variants, CIMORL-TS (Tree Search) and CIMORL-MPPI (MPPI), which leverage privileged global information during training to enable fully decentralized deployment. Experimental validation in cooperative and adversarial scenarios demonstrates a 21% hypervolume improvement and superior policy stability compared to state-of-the-art baselines. Real-world experiments with Crazyflie drones further validate the framework's robustness in resource allocation and multi-attacker multi-defend scenarios under partial observability.@article{marino2026sample, author={Marino, Antonio and Restrepo, Esteban and Chung, Soon-Jo and Giordano, Paolo Robuffo and Pacchierotti, Claudio}, journal={IEEE Transactions on Robotics}, title={Sampling-Based Coordination-Informed Multiobjective Multirobot Reinforcement Learning}, year={2026}, volume={42}, number={}, pages={3570-3589}, keywords={Training;Dynamics;Trajectory;Modeling;Optimization;Vectors;Synchronization;Reinforcement learning;Learning (artificial intelligence);Matrices;Distributed control;graph neural network;multiobjective optimization;reinforcement learning}, doi={10.1109/TRO.2026.3723950}} -
Stable Transformer-Actor-Critic Model Predictive Control : A Contraction Analysis Approach
IEEE Control Systems Letters preprint
Actor-Critic Model Predictive Control (MPC) effectively addresses complex, non-convex control problems, but guaranteeing the closed-loop stability of sequence-based learning models within these pipelines remains challenging. This paper introduces a novel Transformer-Actor-Critic MPC architecture with formal robustness guarantees. First, we prove that Transformer networks can satisfy global incremental Input-to-State Stability. We then leverage Riemannian contraction theory to analyze the interconnected dynamics between the physical plant and the predictive neural network. Finally, we integrate these theoretical bounds as a training regularizer to yield a certifiably robust policy. The framework is validated on a nonlinear 3D drone model executing target-reaching and obstacle-avoidance maneuvers. -
Learning Robot Metareasoning: Adaptive Computation Budgets for Real-Time Control
IEEE International Conference on Robotics and Automation (ICRA) preprint
Planning while moving faces an historical trade-off between deliberation and reactivity. A planner running concurrently with execution needs time to compute, so the robot must commit to an open-loop action chunk for that duration: a longer commitment buys a better plan at the cost of feedback frequency. Existing controllers fix this action commitment, and consequently the computation budget, before deployment. However, the optimal computation time is highly state-dependent (due to factors such as unmodeled dynamics, safety constraints, and marginal optimisation gains), i.e. dynamic, therefore offline static tuning risks leaving systems blind during critical maneuvers or unable to find global effective solutions. We cast the identification of the optimal computation budget as a semi-Markov decision process (SMDP) in which a single decision defines both the robot action commitments and the time allocated to the solver. To solve this, we train a reinforcement-learning policy conditioned on the current state, the executing plan, and inexpensive planner statistics. Across ground-robot navigation, agile quadrotor flight, and contact-rich manipulation, the learned policy changes its budget with the demands of the scenario, discovering efficient reasoning patterns that outperform fixed-budget baselines. -
Weakly Supervised Discovery of Motion Styles for Compositional Robot Control
IEEE International Conference on Robotics and Automation (ICRA) preprint
When a robot possesses more degrees of freedom than a task requires, multiple trajectories can achieve the same objective using distinct motion styles. Compositional robot control aims to preserve a selected style while simultaneously executing an independent task command, effectively balancing these potentially competing objectives. In this paper, we present CODES, a framework for compositional robot control that discovers a coherent style repertoire from diverse recordings and composes it with independent commands. First, motion retargeting maps the recordings to motions executable by the target robot; next, weakly supervised consensus encoding and density-based clustering identify styles within and across recordings without a prescribed count; finally, reinforcement learning trains a policy to compose the discovered styles with independent commands. We evaluate CODES on a Unitree Go2 quadruped across two simulators and physical hardware. Compared to baselines, CODES significantly improves kinematic coherence. Furthermore, our controller successfully tracks velocity commands that generalize beyond the support of the recorded data, exhibiting robust and distinguishable style characteristics during both cross-simulator and real-world hardware deployments. -
Risk-Aware Co-Optimization for Multi-Robot Control under Execution Uncertainty
IEEE Robotics and Automation Letters journal
While multi-robot motion control often treats the environment as a static constraint and assumes near-perfect execution, real-world uncertainty poses significant challenges. In poorly structured or densely populated spaces, these unpredictable variables frequently lead to collision risks and degraded performance. We address this issue by proposing a risk-aware framework that co-optimizes robot actions and environment configurations to mitigate execution uncertainty. To handle the complex relationships among the environment, robots, and uncertainty impact, we develop a bi-level approach using neural surrogate optimization, effectively bridging nonlinear programming with model-free learning for joint optimization. Supported by theoretical robustness guarantees, numerical simulations, and physical experiments, our co-optimization strategy creates ``multi-robot-friendly'' environments that mitigate the impact of execution uncertainty and facilitate superior multi-robot coordination.@article{gao2026risk, author={Gao, Zhan and Marino, Antonio and Prorok, Amanda}, journal={IEEE Robotics and Automation Letters}, title={Risk-Aware Co-Optimization for Multi-Robot Control Under Execution Uncertainty}, year={2026}, volume={}, number={}, pages={1-8}, keywords={Optimization;Robots;Uncertainty;Modeling;Design methodology;Training;Learning (artificial intelligence);Quadrotors;Trajectory;Layout;Multi-robot systems;constrained motion planning;deep learning methods}, doi={10.1109/LRA.2026.3747574}} -
Leave It Better Than Found: Value-Guided Cooperation in Sequential Tasks
Conference on Robot Learning (CoRL) conference
We consider the problem of sequential task continuation, where one robot acts in a workspace that has been affected by a previous robot. Current approaches in multi-robot cooperation rely heavily on co-training or co-optimization, assuming that the robots have access to collective interaction data during the policy training or planning process. However, as increasingly varied robotic embodiments and policies are deployed and operate in the same real-world environment, learning policies for cooperation becomes infeasible due to the combinatorial explosion of possible robot combinations that need to be accounted for. To address this challenge, this paper considers the problem of ad hoc sequential cooperation between robots that have never before operated together in a shared environment, and proposes PAVE (Policy Adaptation via Value Exchange) as a solution. We demonstrate that robots can explicitly communicate their environmental preferences through sharing their value functions, enabling coordination without the need for prior joint training. To ground this exchange and prevent overconfidence in out-of-distribution states, we further introduce a k-nearest neighbors (KNN)-based uncertainty filter. Our method yields an increase of up to 47.7% in task success rates across simulated and real-world experiments compared with purely selfish agents, demonstrating that value function transmission between agents is useful as a basis for ad hoc coordination.
2025
-
A Gated Graph Neural Network Approach to Fast-Convergent Dynamic Average Estimation
ACM Transactions on Intelligent Systems and Technology journal
Dynamic average estimation is a critical problem in multi-agent systems, enabling agents to collaboratively estimate time-varying signals using only local information exchange. Traditional model-based approaches often face challenges related to convergence speed and sensitivity to network topology changes. This article introduces a novel learning-based solution leveraging Gated Graph Neural Networks (GGNNs) for fast-convergent dynamic average estimation in a fully distributed manner. Taking advantage of the inherent structure of GGNNs, the proposed method models the estimation process as a distributed autoregressor, ensuring rapid convergence while maintaining stability. We incorporate a regularization term during training to enforce convergence guarantees and introduce an encoding–decoding mechanism to reduce communication overhead without sacrificing accuracy compared to standard GGNNs. Extensive numerical experiments demonstrate that our approach significantly outperforms conventional model-based estimators in terms of both convergence speed and precision, making it a promising alternative for multi-agent applications that require dynamic average estimation.@article{marino2024average, title={\href{https://dl.acm.org/doi/abs/10.1145/3725857}{A Gated Graph Neural Network Approach to Fast-Convergent Dynamic Average Estimation}}, author={Marino, Antonio and Pacchierotti, Claudio and Robuffo Giordano, Paolo}, journal = {ACM Trans. Intell. Syst. Technol.}, month = may, articleno = {68}, numpages = {18}, year = {2025}, publisher = {Association for Computing Machinery}, address = {New York, NY, USA}, year={2025}, volume = {16}, number = {3}, issn = {2157-6904}, } -
Distributed Multi-Robot Active-Sensing of a Diffusive Source
IEEE Robotics and Automation Letters journal
This letter considers the problem of coordinating a group of mobile robots for distributedly estimating the parameters of a diffusion model that generates a time-varying spatial field. We assume that each robot can measure the local concentration of a substance continuously released in the environment and base the proposed distributed estimation strategy on an Extended Information Consensus Filter (E-ICF) with a forgetting factor. We then develop a decentralized online motion strategy aimed at minimizing a Gramian-based information metric that improves the E-ICF convergence. Additional constraints, among which collision avoidance, are integrated as Control Barrier Functions (CBFs) in a Quadratic Program (QP). Finally, we present statistical comparisons against three baselines which show the improved performance of the proposed method in a range of simulated scenarios, and we also report the results of experiments carried out with quadcopters to demonstrate the actual implementability of the approach and its effectiveness in generating online, collision-free, and informative motions.@article{paganiActive, author={Pagano, Francesca and De Carli, Nicola and Restrepo, Esteban and Marino, Antonio and Giordano, Paolo Robuffo}, journal={IEEE Robotics and Automation Letters}, title={Distributed Multi-Robot Active-Sensing of a Diffusive Source}, year={2025}, volume={10}, number={10}, pages={10807-10814}, keywords={Robots;Robot sensing systems;Estimation;Sensors;Mathematical models;Measurement;Information filters;Observability;Covariance matrices;Convergence;Multi-robot systems;distributed robot systems;active-sensing;gramian}, doi={10.1109/LRA.2025.3606376}, note ={\textbf{\href{https://youtu.be/6zDVlYBqFcE}{video}}},} -
Decentralized Reinforcement Learning for Multi-Agent Multi-Resource Allocation via Dynamic Cluster Agreements
IEEE Robotics and Automation Letters journal
This letter addresses the challenge of allocating heterogeneous resources among multiple agents in a decentralized manner. Our proposed method, Liquid-Graph-Time Clustering-IPPO, builds upon Independent Proximal Policy Optimization (IPPO) by integrating dynamic cluster consensus, a mechanism that allows agents to form and adapt local sub-teams based on resource demands. This decentralized coordination strategy reduces reliance on global information and enhances scalability. We evaluate LGTC-IPPO against standard multi-agent reinforcement learning baselines and a centralized expert solution across a range of team sizes and resource distributions. Experimental results demonstrate that LGTC-IPPO achieves more stable rewards, better coordination, and robust performance even as the number of agents or resource types increases. Additionally, we illustrate how dynamic clustering enables agents to reallocate resources efficiently also for scenarios with discharging resources.@article{marinocluster, author={Marino, Antonio and Restrepo, Esteban and Pacchierotti, Claudio and Giordano, Paolo Robuffo}, journal={IEEE Robotics and Automation Letters}, title={\href{https://hal.science/hal-04972227}{Decentralized Reinforcement Learning for Multi-Agent Multi-Resource Allocation via Dynamic Cluster Agreements}}, year={2025}, volume={10}, number={8}, pages={\\ pp. 8123-8130}, keywords={Resource management;Training;Dynamic scheduling;Reinforcement learning;Scalability;Robot kinematics;Optimization;Decision making;Data mining;Artificial intelligence;Distributed Control;Graph Neural Network;Resource Assignment}, note ={\textbf{\href{https://www.youtube.com/watch?v=lc-TGA1kMiE}{video}}}, } -
Learning to Communicate and Coordinate: Distributed Learning-based Control for Multi-Robot Systems
University of Rennes, Phd Thesis thesis
Multi-robot systems represent a key class of multi-agent systems where multiple autonomous agents cooperate to achieve tasks beyond the capabilities of a single robot. Their effectiveness relies on decentralized or distributed control, where collective behaviors emerge from local interactions under limited information and communication. Classical approaches have enabled significant progress in understanding coordination, information flow, and connectivity maintenance. However, these methods often depend on strong modeling assumptions and struggle to scale in dynamic, uncertain environments. Recent advances in machine learning provide promising alternatives by allowing agents to learn coordination strategies from data, enabling robustness and adaptability under partial observability and noisy sensing. Yet, existing learning-based frameworks rarely account for explicit communication, a critical factor in scalable multi-robot coordination that has been central to control-theoretic approaches. This dissertation addresses this gap by developing hybrid methods that integrate learning with communication-aware distributed control. By combining the generalization and flexibility of machine learning with the theoretical guarantees of control theory, this work advances the foundations of collective decision-making and contributes to the deployment of communication-aware, learning-based multi-robot systems in real-world applications.@phdthesis{marino2025learning, title={Learning to Communicate and Coordinate: Distributed Learning-based Control for Multi-Robot Systems}, author={Marino, Antonio}, year={2025}, school={Universit{\'e} de Rennes} }
2024
-
Workshop on Real-World Challenges in Multi-Robot Cooperation
IROS 2024 workshop
Recent advances in methods, algorithms, computational power, and energetic efficiency have made the deployment of multi-robot systems (MRS) a possibility. Moreover, with the rise of new control methods, learning and perception algorithms, the research community has shown increasing interest in such systems, albeit for relatively simple tasks such as surveillance, monitoring, and warehouse management. For successfully introducing MRS-based in real-world fields, e.g. manufacturing, agriculture and healthcare, there is a need to address more complex and intelligent cooperative tasks. Furthermore, even for simple tasks, there is a big gap between theoretical/lab results and the actual implementation of multi-robot solutions. In light of this, in this workshop, we aim to address questions such as: how do we move from simple tasks to cooperative intelligent missions? What are the new frontiers for MRS? What are the practical and theoretical challenges preventing an increased implementation of MRS-based applications? The objective is to collectively explore novel applications in which MRS could be developed and to identify and tackle the challenges hindering their broader integration in real-world applications. This workshop can hold great value for both the research community and industry, providing a platform to discuss, analyze, and shape the future trajectory of multi-robot systems. We anticipate that the insights gained from this collaborative effort will contribute significantly to advancing the field and addressing the complexities associated with achieving high levels of cooperative behaviours in MRS. -
Graph Neural Network-Based Real-Time 3D Tracking for Micro-Agent Control
MARSS 2024 conference
Micro-surgical robotic systems are gaining prominence in minimally invasive surgery within the medical field. However, accurately tracking the position of the moving agents at the micro-scale remains a significant challenge, particularly for multi-agent systems operating in cluttered and unknown environments. Traditional image analysis methods can falter when confronted with issues such as mutual and obstacle occlusion, especially in dynamic and unstructured scenarios. In order to address this issue, this study introduces a graph-based multi-agent 3D tracking algorithm for a micro-agent control system. This algorithm integrates image information with the control inputs used to navigate the micro agents. We combine the power of Convolutional Neural Networks and Graph Neural Networks to effectively extract features from image sources, and combine them with historical data and control inputs. The primary novelty of this algorithm is its ability to make predictions when the target is occluded in the 2D detection results. The proposed system achieved a tracking error of 0.15 mm, outperforming standard model-based tracking techniques.@inproceedings{jin2024graph, title={\href{https://inria.hal.science/hal-04587941/}{Graph Neural Network-Based Real-Time 3D Tracking for Micro-Agent Control}}, author={Jin, Yuxin and Basualdo, Franco Pi{\~n}an and Marino, Antonio and Mei, Yongfeng and Robuffo Giordano, Paolo and Pacchierotti, Claudio and Misra, Sarthak}, booktitle={International Conference on Manipulation, Automation and Robotics at Small Scales (MARSS)}, year={2024} } -
Liquid-Graph Time-Constant Network for Multi-Agent Systems Control
CDC 2024 conference
In this paper, we propose the Liquid-Graph Time-constant (LGTC) network, a continuous graph neural network (GNN) model for control of multi-agent systems based on the recent Liquid Time Constant (LTC) network. We analyse its stability leveraging contraction analysis and propose a closed-form model that preserves the model contraction rate and does not require solving an ODE at each iteration. Compared to discrete models like Graph Gated Neural Networks (GGNNs), the higher expressivity of the proposed model guarantees remarkable performance while reducing the large amount of communicated variables normally required by GNNs. We evaluate our model on a distributed multi-agent control case study (flocking) taking into account variable communication range and scalability under non-instantaneous communication.@inproceedings{marino2024liquid, title={\href{https://arxiv.org/abs/2404.13982}{Liquid-Graph Time-Constant Network for Multi-Agent Systems Control}}, author={Marino, Antonio and Pacchierotti, Claudio and Robuffo Giordano, Paolo}, booktitle={2024 IEEE 63rd Conference on Decision and Control (CDC)}, year={2024}, volume={}, number={}, pages={793-800}, organization={IEEE} } -
Multi-UAVs End-to-End Distributed Trajectory Generation Over Point Cloud Data
IEEE Robotics and Automation Letters journal
This letter introduces an end-to-end trajectory planning algorithm tailored for multi-UAV systems that generates collision-free trajectories in environments populated with both static and dynamic obstacles, leveraging point cloud data. Our approach consists of a 2-branch neural network fed with sensing and localization data, able to communicate intermediate learned features among the agents. One network branch crafts an initial collision-free trajectory estimate, while the other devises a neural collision constraint for subsequent optimization, ensuring trajectory continuity and adherence to physical actuation limits. Extensive simulations in challenging cluttered environments, involving up to 25 robots and 25% obstacle density, show a collision avoidance success rate in the range of 100−85%. Finally, we introduce a saliency map computation method acting on the point cloud data, offering qualitative insights into our methodology.@article{marinomulti, author={Marino, Antonio and Pacchierotti, Claudio and Robuffo Giordano, Paolo}, journal={IEEE Robotics and Automation Letters}, title={\href{https://ieeexplore.ieee.org/document/10592306}{Multi-UAVs End-to-End Distributed Trajectory Generation Over Point Cloud Data}}, year={2024}, volume={9}, number={9}, note ={\textbf{\href{https://www.youtube.com/watch?v=LSk8gjyuSKA&ab_channel=Rainbow_Inria_Irisa}{video}}}, pages={7629-7636}, keywords={Trajectory;Drones;Safety;Planning;Sensors;Optimization;Heuristic algorithms;Distributed control;graph neural network;trajectory generation}, } -
Input State Stability of Gated Graph Neural Networks
IEEE Transactions on Control of Network Systems journal
In this article, we aim to find the conditions for the input-to-state stability (ISS) and incremental ISS of gated graph neural networks (GGNNs). We show that this recurrent version of graph neural networks can be expressed as a dynamical distributed system and, as a consequence, can be analyzed using model-based techniques to assess its stability and robustness properties. Then, the stability criteria found can be exploited as constraints during the training process to enforce the internal stability of the neural network. Two distributed control examples, i.e., flocking and multirobot motion control, show that using these conditions increases the performance and robustness of the GGNNs.@article{marino2024input, title={\href{https://ieeexplore.ieee.org/document/10458338}{Input State Stability of Gated Graph Neural Networks}}, author={Marino, Antonio and Pacchierotti, Claudio and Robuffo Giordano, Paolo}, journal={IEEE Transactions on Control of Network Systems}, year={2024}, volume={11}, number={4}, pages={2052-2063}, }
2022
-
Gestural and touchscreen interaction for human-robot collaboration: A comparative study
IAS 2022 conference
Close human-robot interaction (HRI), especially in industrial scenarios, has been vastly investigated for the advantages of combining human and robot skills. For an effective HRI, the validity of currently available human-machine communication media or tools should be questioned, and new communication modalities should be explored. This article proposes a modular architecture allowing human operators to interact with robots through different modalities. In particular, we implemented the architecture to handle gestural and touchscreen input, respectively, using a smartwatch and a tablet. Finally, we performed a comparative user experience study between these two modalities.@inproceedings{bongiovanni2022gestural, title={\href{https://link.springer.com/chapter/10.1007/978-3-031-22216-0_9}{Gestural and touchscreen interaction for human-robot collaboration: A comparative study}}, author={ Marino, Antonio and Bongiovanni, Antonino and De Luca, Alessio and Gava, Luna and Grassi, Lucrezia and Lagomarsino, Marta and Lapolla, Marco and Roncagliolo, Patrick and Macci{\`o}, Simone and Carf{\`\i}, Alessandro and others}, booktitle={International Conference on Intelligent Autonomous Systems}, pages={122--138}, year={2022}, organization={Springer} }
2021
-
Reinforcement learning based control for a magnetic flexible endoscope
Hamlyn Symposium on Medical Robotics, 2021 workshop
A reinforcement learning based control of a magnetic actuated endoscope for robotic assisted colonoscopy was presented. Our main aim was to prove that, without the knowledge of the system and the environment, the reinforcement adaptability allows to perform good navigation of the colon beside obstacles and deformations while keeping contact between the endoscope and the colon wall. In fact, this is specifically required in different applications like colonoscopy ultrasound scanning.@inproceedings{marino2021reinforcement, title={Reinforcement learning based control for a magnetic flexible endoscope}, author={Marino, Antonio and Scaglioni, Bruno and Valdastri, Pietro}, booktitle={Proc. Hamlyn Symp. Med. Robot}, pages={1}, year={2021} }
2019
-
PID tuning with neural networks
ACIIDS 2019 conference
In this work we will report our initial investigation of how a neural network architecture could become an efficient tool to model Proportional-Integral-Derivative controller (PID controller). It is well known that neural networks are excellent function approximators, we will then be investigating if a recursive neural networks could be suitable to model and tune PID controllers thus could assist in determining the controller’s proportional, integral, and the derivative gains. A preliminary evaluation is reported.@inproceedings{marino2019pid, title={\href{https://link.springer.com/chapter/10.1007/978-3-030-14799-0_41}{PID tuning with neural networks}}, author={Marino, Antonio and Neri, Filippo}, booktitle={Intelligent Information and Database Systems: 11th Asian Conference, ACIIDS 2019, Yogyakarta, Indonesia, April 8--11, 2019, Proceedings, Part I 11}, pages={476--487}, year={2019}, organization={Springer} }