Research

Multi-Agent Learning and Coordination — Theory and Applications

Across our work on multi-agent systems, a unifying theme is the development of scalable, decentralized frameworks for long-horizon coordination under uncertainty, bridging rigorous theory with multi-robot system design. On the theoretical side, our research emphasizes distributed decision-making mechanisms—rooted in control, game theory, and learning—that enable multiple agents to coordinate using local information while achieving globally consistent objectives, with guarantees on stability, efficiency, and convergence. These frameworks explicitly address fundamental challenges such as limited communication, partial observability, and the trade-off between coupling (coordination) and decoupling (independent operation), which are critical for scalability in large teams. Complementing this, our application-driven work demonstrates how these principles translate into real-world tasks, including cooperative manipulation, task allocations, and multi-target tracking, where coordination must be robust to model uncertainty and execution mismatch. A key cross-cutting idea is the integration of learning with structured models, enabling adaptive coordination policies that evolve from data while retaining interpretability and reliability. Together, these efforts establish a cohesive framework for designing multi-agent systems that are not only theoretically grounded but also capable of sustained, high-performance operation in complex physical environments.

Read More:

• D. Alvear, G. Turkiyyah, and S. Park, "Cooperative Grasping for Collective Object Transport in Constrained Environments," IEEE Robotics and Automation Letters, Jan. 2026 (link, video).

• L. C. D. Bezerra, A. M. G. dos Santos, and S. Park, "Learning Policies for Dynamic Coalition Formation in Multi-Robot Task Allocation," IEEE Robotics and Automation Letters, July 2025 (link, video).

• S. Park and N. E. Leonard, "Learning with Delayed Payoffs in Population Games using Kullback-Leibler Divergence Regularization," IEEE Transactions to Automatic Control, April 2025 (link).

• J. Chen, M. Abugurain, P. Dames, and S. Park, "Distributed Multi-Robot Multi-Target Tracking Using Heterogeneous Limited-Range Sensors," IEEE Transactions on Robotics, Feb. 2025 (link).

Long-Horizon Task and Sequence Planning for Robotic Manipulation

With rapid advancements in sensing, actuation, and mechanical design, robotic systems are increasingly capable of executing complex, high-precision operations across diverse domains—from autonomous driving in urban environments to large-scale package handling in fulfillment centers. In industrial settings, however, many tasks are inherently long-horizon and sequential, requiring robots to plan and execute extended sequences of interdependent manipulation actions under physical constraints.

Achieving reliable long-horizon sequence planning in real-world environments critically depends on accurately modeling how robot actions affect physical objects over time. Traditional robotics approaches rely on analytically tractable mathematical models, which offer interpretability and theoretical guarantees. However, as manipulation tasks grow in complexity—particularly in contact-rich and unstructured environments—such models become increasingly difficult to derive and insufficient to capture real-world dynamics.

To address this gap, we propose a computational modeling framework that leverages physics-based simulation as a core component of long-horizon planning. Specifically, we construct high-fidelity simulation models using physics engines and develop feedback-driven algorithms to continuously adapt model parameters (e.g., mass, inertia, friction) based on discrepancies between simulated predictions and real-world observations. This adaptive simulation enables more reliable prediction of action outcomes over extended planning horizons.

Building on this foundation, we aim to develop long-horizon sequence planning algorithms for robotic manipulation, where policies are trained and validated within these adaptive simulation environments before deployment. Our primary application focus is on industrial automation tasks—such as assembly operations in manufacturing lines, laboratory workflows, and precision agriculture—that are labor-intensive, sequential in nature, and require high levels of dexterity and precision.

Read More:

• C. Shu, A. Kim, and S. Park, "Subassembly to Full Assembly: Effective Assembly Sequence Planning through Graph-based Reinforcement Learning," 2025 International Conference on Robotics and Automation (link, video).

Networked Robots for Monitoring Underwater Habitats

Coral reefs, surrounded by vibrant marine life, constitute rich and diverse habitats. Given their ecological and economic significance, continuous monitoring of their key areas is paramount for marine biology research, economic development, and environmental protection. The extended survey and observation of vast marine environments are labor-intensive, necessitating technological advancements. While many efforts have been made in research and development to create autonomous robots capable of surveying and collecting data in remote areas inaccessible to human divers or requiring long-term and repeated monitoring, technical challenges remain. These challenges include implementing navigation strategies for individual underwater robots to operate in close proximity to marine life for extended periods while ensuring safety and non-interference with natural habitats. Additionally, the severe signal attenuation underwater poses technical obstacles in information exchange and achieving a high degree of coordination among multiple robots tasked with surveying large-scale coral reef environments. The project explores learning-based planning and control approaches for individual robots to acquire safe navigation strategies and adaptively apply them in response to environmental conditions, such as light conditions and current speeds. Furthermore, drawing from theoretical studies in multi-agent coordination, we aim to investigate how the robots can adopt optimal team strategies while working under the constraints of limited information exchange.

Multi-Robot Perception, Manipulation, and Learning

An important question in multi-agent system research is "how to synergize individual agents’ capabilities in sensing, actuation, and communication?" This research aims at discovering an effective multi-agent decision-making model for a robotic team to learn and attain optimal team strategies in dynamic and resource-constrained environments. We apply outcomes of the research to conceiving a heterogenous robotic team that can be used to address real societal problems such as waste collection/removal.

Read More:

• S. Park, Y. Desmond Zhong, and N. E. Leonard, "Multi-Robot Task Allocation Games in Dynamically Changing Environments, " IEEE International Conference on Robotics and Automation (ICRA), June 2021 (link).

Social Trajectory Planning for Autonomous Surface Vessels

We imagine, in the future, a fleet of autonomous vessels to be employed as a new means of transportation through city's waterways. To ensure safety in vessel navigation, we need to understand the navigation behaviors of human-operated vessels and allow autonomous vessels to learn those behaviors. We propose a new approach called "Social Trajectory Planning" to address the key challenge in guaranteeing safety in canal navigation. Through validation using vessel trajectory data set, we demonstrate that Social Trajectory Planning reduces the chances of vessel-to-vessel collision and improves the predictability of autonomous vessels' movement.

Read More:

• S. Park, M. Cap, J. Alonso-Mora, C. Ratti, and D. Rus, “Social Trajectory Planning for Urban Autonomous Surface Vessels,” IEEE Transactions on Robotics, Oct. 2020 (link).

Higher-Order Learning in Population Games

This work formalizes decision-making models and proposes analytical methods to establish stability in strategic interactions among a vast number of decision-making agents. In this context, evolutionary dynamics models describe how the portions of agents adopting each available strategy evolve in response to the payoff ascribed by a payoff dynamics model. Motivated by its importance in developing effective multi-agent decision-making algorithms in a wide range of engineering applications, this work describes a new class of payoff dynamics models and analytical methods that hinge on passivity-based techniques to characterize the stability of evolutionary dynamics model under payoff dynamics models.

Read More:

• S. Park, J. S. Shamma, and N. C. Martins, "From Population Games to Payoff Dynamics Models: A Passivity-Based Approach," IEEE Conference on Decision and Control (CDC), Dec. 2019 (tutorial session).

• S. Park, N. C. Martins, and J. S. Shamma, "Payoff Dynamics Model and Evolutionary Dynamics Model: Feedback and Convergence to Equilibria," arXiv:1903.02018.

Roboat: Fleet of Reconfigurable Robotic Vessels 

At the intersection of urban mobility and self-reconfigurable modular robotics, this research project seeks a new approach for design and control of a fleet of autonomous surface vessels (ASV) that can navigate and self-assemble into dynamic infrastructures in city's waterway. We conceive a heterogeneous team of reconfigurable ASVs and develop multi-vessel control and planning algorithms that enable the fleet's core capabilities for navigation and self-assembly in canal environments. We envision this fleet of autonomous vessels will serve as a new means of transportation and on-demand infrastructure in the near future. Check out Roboat website for more.

Read More:

• B. Gheneti, S. Park, R. Kelly, D. Meyers, P. Leoni, C. Ratti, and D. Rus, "Trajectory Planning for the Shapeshifting of Autonomous Surface Vessels," 2nd IEEE International Symposium on Multi-Robot and Multi-Agent Systems (MRS' 19), Aug. 2019.

• S. Park, E. Kayacan, C. Ratti, and D. Rus, "Coordinated Control of a Reconfigurable Multi-Vessel Platform: Robust Control Approach", IEEE International Conference on Robotics and Automation (ICRA), May 2019.

Underworlds: Smart Sewage Infrastructure for Urban Epidemiology 

Underworlds proposes to develop a human health census by sampling the “urban gut” through collecting and analyzing sewage samples at high-level of spatial and temporal resolutions. We design a smart sewage infrastructure consisting of networked sensing/sampling devices, a computational approach to collect most informative sewage samples (subject to budget constraints), and data analysis/visualization pipelines. Our team has been successfully deployed Underworlds around the globe. Check out Underworlds website for more. 

Read More:

• E. D Evans, C. Dai, S. Isazadeh, S. Park, C. Ratti, and E. J. Alm, "Longitudinal wastewater sampling in buildings reveals temporal dynamics of metabolites," PLOS Computational Biology, June 2020 (link).

• S. Park, C. Ratti, and D. Rus, "Adaptive Sensor Selection for Monitoring Stochastic Processes," IEEE Conference on Decision and Control (CDC), Dec. 2018.

Animal-borne Wireless Camera Network: Remote Imaging of Community Ecology 

This collaborative research aims at designing, constructing, and field-testing a standalone networked animal‐borne monitoring system conceived to study community ecology remotely. The system is designed to use information exchange across the network for the devices to jointly decide without supervision when and how to use each sensing modality. The significance of the system is in its capability to be programmed to selectively document events of interactions in animal groups and its ability to operate for an extended period of time in natural habitats of study animals. Through system deployment in Gorongosa National Park (Mozambique), our team validated its long-term data collection capability and collected animal behavioral data which have been used as important resources in ecology studies.

Read More:

• S. Park, K. H. Aschenbach, M. Ahmed, W. Scott, N. E. Leonard, K. Abernathy, G. Marshall, M. Shepard, and N. C. Martins, “Animal-Borne Wireless Network: Remote Imaging of Community Ecology,” Journal of Field Robotics, July 2019 (link, multimedia).

Distributed State Estimation for Omniscience

Motivated by challenges in estimation over a network of sensing devices and its importance in engineering applications such as connected-vehicle control and environmental monitoring, this work proposes a distributed estimation algorithm for a network of agents to estimate a state vector of a large-scale dynamical system over a communication-constrained network. The key findings of the work are that we characterize exact conditions under which every agent asymptotically recovers the system’s entire state vector (with exponential recovery rate) – state omniscience – and the complexity of the algorithm depends linearly on the size of the system – scalability. The work becomes a central piece in designing a distributed sensor fusion algorithm for the animal-borne wireless camera network to effectively document group behavior of wild animals.

Read More:

• S. Park and N. C. Martins, "Design of Distributed LTI Observers for State Omniscience," IEEE Transactions on Automatic Control, Vol.62, No.2, Feb. 2017 (link).