⚡ electrical engineering

Field-Validated Formation Control for Autonomous Surface Vehicle Fleets in Aquatic Environments with Currents

This paper presents and validates a novel navigation function-based formation control method for autonomous surface vehicle fleets that enables them to maintain relative positioning while drifting with currents, successfully demonstrating robust performance in field experiments involving GPS uncertainty, communication dropouts, and dynamic aquatic environments.

Original authors: Chanaka Thushitha Bandara, Herbert G. Tanner

Published 2026-09-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chanaka Thushitha Bandara, Herbert G. Tanner

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Rivers, lakes, and coastal estuaries are under increasing pressure from human activity and climate change, creating an urgent need to monitor water quality across vast, shifting areas. Fixed monitoring stations can only watch a single spot, but fleets of small, robotic boats equipped with sensors can cover far more ground. The challenge lies in getting these robots to work together effectively. They must maintain a specific shape relative to one another to gather accurate data, yet they operate in a chaotic environment filled with strong currents, floating obstacles, and unreliable wireless signals. If the robots fight the current to hold a fixed position on a map, they waste precious energy. If they drift too freely, they lose the formation needed to reconstruct a clear picture of the water's condition. The goal is to find a way for a group of robots to stay in formation while letting the river carry them along, all while ensuring they never crash into each other or the shore.

Researchers at the University of Delaware have developed and tested a new control method that allows fleets of autonomous surface vehicles to do exactly this. Instead of forcing the robots to hold a rigid position in space, the system guides them to maintain their relative spacing to one another, allowing the entire group to drift with the current like a school of fish moving with a stream. This approach was validated through computer simulations and real-world field tests using four small, commercially available robotic boats called Jaiabots. The team first tested the system in the calm, controlled waters of Lake Allure in Pennsylvania, and then moved to the much more challenging environment of the Delaware River in Philadelphia, where the robots had to contend with tidal currents, GPS signal errors, and intermittent communication dropouts.

The core of the solution is a mathematical framework that acts like a set of invisible rules for the robots. Imagine the water surface as a landscape where the desired formation is a valley and obstacles are hills. The robots are programmed to naturally roll down toward the valley while avoiding the hills. However, standard versions of this "potential field" method had two major flaws for river use: they did not account for the physical limits of the boats, such as how tightly they can turn or how slowly they can move, and they did not prevent the robots from drifting into the riverbanks. Furthermore, classical potential field methods often suffer from "local minima," which are like shallow dips in the landscape where a robot can get stuck, unable to reach its true destination. To fix this, the researchers specially constructed this mathematical landscape so that there are no local minima where the robots could get trapped away from their desired position. They also added "virtual" robots that act as guardians for the river boundaries, ensuring the real boats stay within safe limits. They also smoothed out the control signals to prevent the rudders from jittering, which can damage the hardware. Finally, they added a forward-looking strategy that plans a short path ahead of time, allowing the boat to navigate around its own turning limitations even when the ideal path seems impossible.

In the lake experiments, the four robots successfully formed a straight line at a 45-degree angle, maintaining a spacing of 10 meters between them. They moved smoothly, avoiding a buoy placed in the water and staying clear of the shoreline, even as wind and minor water movement pushed them around. The system proved robust enough to handle the fact that real boats cannot stop instantly or turn on a dime; when the ideal path was blocked by a turning constraint, the robots found a feasible detour and corrected their course in the next moment. The data showed that the distance between any two robots never dropped to zero, confirming that collisions were successfully avoided, and the group consistently converged on the desired shape.

The true test came in the Delaware River, where the conditions were far less forgiving. The river is approximately 2,250 feet wide, with an average current speed of 1.3 feet per second. The team deployed the robots from a pier, letting them drift and maneuver into a line formation that stretched across a significant portion of the river. During one trial, the formation was set to span the full width of the river, with the robots spaced about 400 feet apart. This setup pushed the limits of the wireless communication range. At one point, the robot furthest from the shore lost contact with the central control hub for several seconds. The system handled this gracefully: the disconnected robot simply stopped moving and waited, while the others continued to adjust their positions. Once the connection was restored, the robot resumed its place, and the entire fleet converged into the correct formation without crashing or getting lost.

In a second river trial, the team adjusted the formation to be tighter, reducing the spacing to about 200 feet to keep all robots within a more reliable communication range. This change resulted in smoother trajectories and no communication losses, demonstrating that the system is flexible enough to adapt to different operational needs simply by changing the desired spacing between the boats. Throughout both river tests, the robots successfully navigated around other watercraft and maintained their formation despite the unpredictable push of the tide and the occasional loss of GPS precision. The results confirm that it is possible to control a fleet of autonomous boats in a dynamic, confined river environment without fighting the current, provided the control system accounts for the physical reality of the machines and the limitations of their communication.

This work bridges the gap between theoretical control algorithms and the messy reality of field robotics. By proving that these methods work on actual hardware in a real river, the researchers have shown a viable path for deploying sensor fleets to monitor water quality in estuaries and coastal zones. The ability to let the fleet drift with the current while maintaining a precise relative shape means these robotic groups can operate longer and cover more ground with less energy. Future work will focus on removing the need for a central control hub, allowing the robots to communicate directly with one another, which would make the system even more robust for open-water missions. For now, the successful trials in the Delaware River stand as a concrete demonstration that autonomous fleets can be trusted to work together in the wild, unpredictable waters of the natural world.

Technical Summary: Field-Validated Formation Control for Autonomous Surface Vehicle Fleets in Aquatic Environments with Currents

Problem Statement
The paper addresses the challenge of controlling fleets of Autonomous Surface Vehicles (ASVs) in confined aquatic environments characterized by spatially varying currents, obstacles, and workspace boundaries. The primary objective is to achieve a formation defined by desired fixed relative positions among agents while allowing the entire fleet to drift with the ambient current, rather than expending energy to maintain absolute global positions. This approach is critical for applications such as monitoring salinity gradients in estuaries, where sensor placement relative to one another is more important than absolute location.

The control design must satisfy several constraints:

Collision Avoidance: Preventing collisions between agents and with environmental boundaries (e.g., shorelines, moored boats).
Kinematic Constraints: Accounting for nonholonomic vehicle dynamics, specifically minimum turning radii and minimum forward speeds (actuator dead zones) that prevent arbitrary slow movement.
Robustness: Handling GPS uncertainty, intermittent communication delays, and external disturbances.
Energy Efficiency: Avoiding the need for agents to fight the current for station-keeping.

Methodology
The authors extend a multi-agent navigation function framework to address the specific limitations of existing methods in riverine environments. The core methodology involves three technical innovations:

Differentiable Potential Function Approximation:
The original navigation function formulation relies on a non-smooth minimum distance function, which can cause control chattering and is difficult to implement on physical hardware with actuator time constants. The authors introduce a conservative, differentiable approximation of the minimum distance function using a log-sum-exp formulation. This eliminates chattering while preserving the collision avoidance guarantees of the original potential field.
Virtual Agents for Boundary Avoidance:
To handle workspace boundaries and internal obstacles (which the original formulation did not address), the authors introduce $n$ additional "virtual" agents. Each real ASV $i$ is paired with a virtual agent $i+n$ positioned at the point of minimum distance between the real robot and the nearest boundary or obstacle. These virtual agents are integrated into the multi-agent potential field, effectively repelling the real robots from boundaries while maintaining the relative formation geometry.
Receding Horizon Waypoint Strategy:
To accommodate kinematic constraints (minimum turning radius and minimum surge speed), the authors employ a receding horizon control layer. Instead of commanding velocities directly along the negated gradient of the navigation function (which may be kinematically infeasible), the system integrates the gradient over a prediction horizon $T$ to generate a local waypoint. The vehicle then executes feasible control inputs to reach this waypoint. This approach allows the fleet to navigate around kinematic limitations and actuator dead zones, ensuring the navigation function value decreases over time even if the vehicle cannot follow the exact flow lines of the potential field.

Key Contributions
The paper claims three specific contributions that bridge the gap between theoretical navigation functions and field deployment:

A differentiable approximation of the potential function that eliminates control chattering, making the method suitable for real-world rudder actuators.
A mechanism for workspace boundary and obstacle avoidance using virtual agents, extending the collision avoidance guarantees to include environmental constraints.
A receding horizon navigation layer that accommodates constrained vehicle kinematics, including minimum speed and turning radius limits, which are critical for underactuated ASVs.

Experimental Validation and Results
The methodology was validated through numerical simulations and two distinct field experiments using four Jaiabot micro-ASVs.

Numerical Simulation: Simulations with point robots confirmed that the proposed controller achieves convergence to the desired formation without collisions, outperforming a baseline Jacobi-coordinate controller which failed to prevent inter-agent collisions.
Lake Environment (Lake Allure, PA): In a controlled setting with minimal currents, the fleet successfully converged to a 45-degree line formation with 10m spacing. The system handled GPS errors and actuator dynamics, maintaining strictly positive inter-robot distances and avoiding the elliptical boundary and a buoy obstacle.
River Environment (Delaware River, Penn's Landing, PA): Experiments were conducted under tidal currents (approx. 1.3 ft/s) with a formation spanning a significant portion of the river width.
- Robustness: The fleet converged to the desired formation despite GPS uncertainty, the presence of other watercraft, and intermittent communication losses. In one trial, communication with a robot 1300 ft away was lost; the robot paused until reconnection, after which the fleet successfully resumed convergence.
- Flexibility: The formation spacing was adjusted mid-experiment (from ~400 ft to ~200 ft) simply by modifying the desired relative-position vector $c$ , demonstrating the ability to balance spatial coverage with communication reliability without altering the control architecture.
- Drift: The fleet successfully maintained relative positioning while drifting with the current, avoiding the energy cost of fighting the flow.

Significance and Claims
The paper asserts that this work represents a significant step toward practical field deployment of formation control for ASVs. Unlike many existing methods that are limited to simulation or assume ideal communication and open spaces, this approach has been validated in "shallow, confined waterways with obstacles" under realistic conditions.

The authors emphasize that the system allows the fleet to "drift with the current rather than spending energy to hold fixed positions," a crucial efficiency gain for long-duration monitoring tasks. Furthermore, the integration of virtual agents and receding horizon control provides a provably convergent, collision-free framework that handles the nonholonomic constraints and actuator limitations inherent to commercial micro-ASVs. The successful recovery from communication dropouts and the ability to adapt formation geometry on the fly highlight the robustness required for real-world environmental monitoring.

Technical Summary: Field-Validated Formation Control for Autonomous Surface Vehicle Fleets in Aquatic Environments with Currents

More like this