Research

We focus on AI-augmented decentralized frameworks for multi-agent environmental mapping, particularly in scenarios with limited sensing and communication. While traditional coverage strategies work well with accurate prior maps, their performance degrades under uncertainty or biased information. Our approach adaptively refines local density estimates using density-driven optiamal control (D²OC), an optimal transport–based framework, combined with a dual multilayer perception (MLP) module that infers local statistics and guides exploration of rarely visited areas.

This self-correcting mechanism ensures scalability, robustness, and convergence under the Wasserstein metric. It can be applied to a variety of environmental monitoring tasks, including detecting landfill gas distributions, where complex, multi-modal spatial patterns must be accurately reconstructed despite sensing and communication constraints. Simulations show our framework achieves significantly higher fidelity than conventional methods.

Publication:
arXiv:  available soon.
Github: available soon.

Simulation results for multi-agent mapping of landfill gas distributions:  (a) Initial gas distribution estimation (blue), unknown ground-truth distribution (shaded red), and starting positions of five linearized quadcopter systems (orange squares); (b) Sample distribution after mission completion without AI augmentation; (c) AI-augmented D²OC results, showing improved coverage compared to (b).

We study multi-agent systems for area coverage in applications such as search-and-rescue, environmental monitoring, and precision agriculture. Unlike uniform coverage, non-uniform coverage prioritizes certain regions, requiring coordination under dynamic and communication constraints. While existing density-driven methods distribute agents according to a reference density, they often fail to guarantee connectivity, which can reduce coverage quality in practice.

Our work introduces a connectivity-preserving approach within the Density-Driven Optimal Control (D²OC) framework. The coverage problem is formulated as a quadratic program using the Wasserstein distance between agent distributions and a reference density. Communication constraints are handled through a smooth penalty function, ensuring global optimality while naturally maintaining inter-agent connectivity without enforcing rigid formations.

Simulations show that our method keeps agents within communication range, enhancing coverage quality and convergence speed compared with conventional approaches that do not explicitly enforce connectivity.

Publication:
• IEEE Control Systems Letters (L-CSS)
DOI: available soon.
arXiv: https://arxiv.org/abs/2511.18579

(a) Multi-agent trajectories without (b) Multi-agent trajectories with (c) Inter-agent distances without (d) inter-agent distances with
connectivity constraint connectivity constraint connectivity constraint connectivity constraint

Modern large-scale farms require precise and sustainable strategies for managing pests, weeds, and crop diseases. Conventional blanket spraying methods often waste chemicals and harm the environment, while single-UAV solutions suffer from limited battery life and payload capacity.

To address these challenges, we develop a Density-Driven Optimal Control (D²OC) framework that enables adaptive, large-area pesticide spraying using multiple UAVs. The method leverages Optimal Transport (OT) theory to allocate spraying effort based on real infestation severity rather than uniform coverage, significantly reducing unnecessary chemical use.

Our UAVs are modeled as linear time-varying systems to account for changing payload mass during spraying, and the D²OC controller, derived through Lagrangian mechanics, coordinates the fleet to achieve:

• Adaptive, priority-aware spraying

• Balanced workload across UAVs

• Reduced mission time and chemical consumption

• Scalable multi-UAV operations for large farms
  
  

Simulation studies show that D²OC outperforms traditional uniform spraying and methods such as Spectral Multiscale Coverage (SMC) in both efficiency and sustainability, offering a powerful solution for next-generation precision agriculture.

Publication:
• IEEE Transactions on Control Systems Technology (TCST)
DOI: https://doi.org/10.1109/TCST.2025.3631091
arXiv: https://arxiv.org/abs/2511.12492

[Three-Quadcopter Trajectories and Residual Weed Density Colormap: (Left) Lawn-Mower Method; (Middle) Spectral Multiscale Coverage; (Right) D²OC. ]

Geodesic path planning is crucial in applications such as robotics, computer graphics, and autonomous navigation, focusing on finding the shortest path between two points on a curved surface while accounting for intrinsic geometry. Traditional methods, including energy function minimization, heat flow, and curvature-based techniques, often face local minima and computational inefficiencies, particularly in irregular environments. This study investigates a novel method that integrates the exponential map with the injectivity radius, ensuring globally optimal paths. Our approach avoids local minima, guarantees the shortest path, and provides real-time computational efficiency. Simulations show that our method outperforms existing techniques in optimality, local minima avoidance, and computation time.

                            [Geodesic Path on the Simple Manifold] [Geodesic Path on the Simple Manifold]

This study addresses a wildlife monitoring problem using a team of UAVs for efficient monitoring of wildlife. The state-of-the-art technology using UAVs has been an increasingly popular tool to monitor wildlife compared to the traditional methods such as satellite imagery-based sensing or GPS trackers. However, there still exist unsolved problems as to how the UAVs need to cover a spacious domain to detect animals as many as possible. In this research, we propose the optimal transport-based wildlife monitoring strategy for a multi-UAV system, to prioritize monitoring areas while incorporating complementary information such as GPS trackers and satellite-based sensing. Through the proposed scheme, the UAVs can explore the large-size domain effectively and collaboratively with a given priority. The time-varying nature of wildlife due to their movements is modeled as a stochastic process, which is included in the proposed work to reflect the spatio-temporal evolution of their position estimation. In this way, the proposed monitoring plan can lead to efficient wildlife monitoring with a high detection rate. Various simulation results including statistical data are provided to validate the proposed work.

[Schematic of Wildlife Monitoring using a Multi-UAV system] [Simulation Results for Wildlife Detection with Three UAVs]

Cave and karst surveying is challenging because surveyors are at risk of danger when exposed to extreme environments (cool air, water, and rock) for a long period. Another risk is that surveyors may get lost or trapped in a cave. Caves and karsts are particularly challenging to survey because of the small size as well as the maze-like nature of the passages. For this reason, roboticists have developed special types of robots in efforts to replace human surveyors with robots. Especially, flying robots (drones) are known to be superior to other types of robots as they have the ability to overcome extreme environments. However, drones are quite limited in terms of their operation time (most of the commercial drones can only fly up to 25 minutes). This is a critical problem as many caves and karsts have lengthy passages that cannot be covered in a short time.

In this project, we aim to develop a helium-assisted hybrid drone for sustainable cave and karst research. Although few drones can fly more than 30 minutes, these drones are classified as medium/large drones, which is not appropriate for cave and karst explorations due to their size. Considering these factors, the objective of this research is to 1) increase the drone flight time at least by an hour; 2) maintain the size of the drone as small as possible. The funding from the NCKRI-NMT Internal Seed Grant Program supports this project to achieve the given project goal, which will eventually serve as a pathway to realize sustainable cave and karst surveys using a hybrid drone.

This research addresses the efficient exploration problem based on the optimal transport theory. The efficiency in this context implies how a team of heterogeneous robots (or agents) covers areas of interest intelligently. An information density distribution that describes the relative importance or priority of regions in the domain is associated with the problem formulation for efficient multi-robot explorations. This scheme is expected to be applicable to wide areas such as weather monitoring, search and rescue, military mission, surveillance and inspection, wildlife monitoring, and planetary exploration. In this study, a new method based on the optimal transport theory is proposed to yield exploration efficiency. The optimal transport theory that quantifies the distance between two probability density functions is employed as a tool to measure as well as to realize efficient multi-robot exploration.

[Single-Robot Exploration] [Multi-Robot Exploration]

• Application to NASA Mars 2020 rover mission: The bottom figures present the site on Mars (Jezero Crater) for the NASA Mars 2020 rover mission. The density distribution is described for areas of interest with a given priority to seek signs of ancient habitable conditions and past microbial life. Instead of sending a single rover, a team of rovers can be deployed to increase the chance to find any signs of past life. The right figure presents the simulation result based on the developed efficient/collaborative multi-rover exploration scheme.

In this study, a novel snake-like robot design is presented and analyzed. The structure described desires to obtain a robot that is most like a snake found in nature. This is achieved with the combination of both rigid and soft link structures by implementing a 3D printed rigid link and a soft cast silicone skin. The proposed structure serves to have a few mechanical improvements while maintaining the positives of previous designs. The implementation of the silicone skin presents the opportunity to use synthetic scales and directional friction. The design modifications of this novel design are analyzed on the fronts of the kinematics and minimizing power loss. Minimization of power loss is done through a numerical minimization of three separate parameters with the smallest positive power loss being used. This results in the minimal power loss per unit distance. This research found that the novel structure presented can be effectively described and modeled, such that they could be applied to a constructed model.

This research addresses an energy-balanced leader-switching policy for formation rotation control of multi-agent systems inspired by bird flocks. Birds that flock in V-formation with a leader rotation strategy are able to travel longer distances due to reduced drag and therefore less energy expenditure. This flocking behavior with a leader rotation will result in more conservation of overall energy and will be particularly beneficial to migrating birds that should fly long distances without landing.  In this study, we propose an energy-balanced leader-switching policy inspired by this bird flocking behavior in order to increase the flight range for multi-agent systems. The formation control of multi-agent systems is achieved by the consensus algorithm, which is fully decentralized through the use of information exchanges between agents. 

This work investigates the utilization of the IR (infrared) camera for outdoor target positioning. The Wii remote IR camera is selected as a platform, which has been widely used in various applications for the following reasons: detection of up to four IR light sources with a fast frame rate (100Hz) and a relatively low price. However, previous Wii remote IR camera applications are limited to indoor uses due to the obvious reason - sunlight interference for outdoor applications. In this research, a signal modulation technique is introduced, which enables the IR camera to look for a particular pattern encoded in an IR beacon. In this way, the IR camera can distinguish the IR beacon from the sunlight interference. The irradiance of the sunlight reflection is also analyzed to guarantee that the IR camera can detect the IR beacon even under extremely sunny weather conditions. As the Wii remote IR camera sensor is overloaded under an extremely bright condition that blocks the camera to see any light sources, we propose the use of a filter to dim the camera. Experimental results for outdoor tests are provided to validate the proposed methods.

This study proposes a receding-horizon, multi-objective optimization approach for robot motion planning in disaster response scenarios. During a search and rescue mission, a robot is deployed in the disaster area to find and egress all victims. In doing so, multiple criteria characterize the effectiveness of such plan. We define three objective functions (performance, uncertainty about victim locations, and uncertainty about the environment) and formulate a multi-objective optimization problem employing a combined weighted-sum and  e-constraint method. To handle dynamic scenarios, we employ a receding-horizon approach that allows to dynamically adapt the constraint.

This study investigates the performance and robustness analysis of a distributed networked control system (DNCS) that has random communication delays between multiple subsystems (agents). To deal with the stability analysis for such DNCSs with communication delays, we adopt a Markov jump linear system framework. Compared to the current state-of-the-art that only guarantees asymptotic stability, our contribution is to develop a unifying framework by adopting an optimal transport theory, which enables both transient and asymptotic performance analysis without assuming any structure (e.g. Markov) on the underlying jump process. 

In the near future, massively parallel computing systems will be necessary to solve computation intensive applications such as multi-physics multi-scale simulations of natural and engineering systems. The key bottleneck in massively parallel implementation of numerical algorithms is the synchronization of data across processing elements (PEs) after each iteration, which results in significant idle time. Thus, there is a trend towards relaxing the synchronization and adopting an asynchronous model of computation to reduce idle time. However, it is not clear what is the effect of this relaxation on the stability and accuracy of the numerical algorithm. In this research we develop a new method to analyze the stability, convergence rate, and probability of error for the asynchronous parallel numerical algorithm by employing the switched dynamical system framework.

Odometry using wheel encoders provides fundamental pose estimates for wheeled mobile robots. Systematic errors of odometry can be reduced by the calibration of kinematic parameters. In this research, an accurate calibration scheme of kinematic parameters is proposed. The contributions of this paper can be summarized as two issues. The first contribution is to present new calibration equations that remarkably reduce the systematic error of odometry. The new equations were derived to overcome the limitation of the conventional schemes. The second contribution is to propose the design guideline of the test track for calibration experiments. The calibration performance can be significantly improved by appropriate design of the test track. 

Control problems of a car-like vehicle are not easy because of nonholonomic velocity constraints. This research proposes a parking control strategy which is composed of an open loop path planner and a feedback tracking controller. By employing a trajectory tracking controller for a two wheeled robot, a car-like vehicle can be successfully controlled to the desired configuration. Experimental results with a radio controlled model car clearly show that the proposed control scheme is practically useful.