Publications

Abstract Unmanned Aerial Vehicles (UAVs) play an increasingly vital role in wilderness Search and Rescue (SAR) operations by enhancing situational awareness and extending the capabilities of human teams. Yet, a lack of standardized benchmarks has impeded the systematic evaluation of single- and multi-agent path-planning algorithms. This paper introduces an open-source dataset and evaluation framework to address this gap. The framework comprises 60 geospatial scenarios across four distinct European environments, featuring high-resolution probability maps. We present a lost person probabilistic model derived from statistical models of lost person behavior. We provide a suite of tools for evaluating search paths against five baseline methods — Spiral, Concentric Circles, Pizza Zigzag, Greedy, and Random Exploration — using three quantitative metrics: Accumulated Probability of Detection, Time-Discounted Probability of Detection, and Lost Person Discovery Score. By providing a structured and extensible framework, this work establishes a foundation for the rigorous and reproducible assessment of UAV search strategies in complex wilderness environments.

Supported by Innovation Fund Denmark (DIREC, 9142-00001B), Independent Research Fund Denmark (NAMUR, 10.46540/4264-00105B), and the WildDrone MSCA Doctoral Network, EU Horizon Europe grant 101071224.

Abstract Multi-UAV systems typically require complex infrastructure to deploy in real-world scenarios, limiting their accessibility and scalability. In addition, current research often relies on custom solutions or proprietary hardware to facilitate inter-UAV communication. In this paper, we propose an open-source, low-cost, plug-and-play solution to enable decentralized UAV-to-UAV communication over 2.4 GHz Wi-Fi using a connectionless protocol. Our approach simplifies the deployment of decentralized systems by allowing UAVs to easily exchange any type of binary data, seamlessly interfacing with ROS 2. The solution uses an ad-hoc style network that allows UAVs to join or leave dynamically without requiring centralized governance or a priori configuration. We describe the architecture of the system, assess the network performance in an outdoor environment using UAVs, and evaluate the system's ability to share information as a swarm through hardware-in-the-loop (HITL) experiments. HITL experiments show that a decentralized planning algorithm running on three simulated UAVs can effectively reach consensus on decentralized task allocation.

Abstract Drones are currently an indispensable tool for emergency response teams performing wilderness Search And Rescue (SAR), as they can cover large and possibly inaccessible areas efficiently. It is, however, still unclear how a drone operator can effectively engage and control a system composed of multiple autonomous robots, especially in unstructured and outdoor environments. This paper reports on ongoing work in the project HERD — Human-AI Collaboration: Engaging and Controlling Swarms of Robots and Drones — in which we focus on how to enable an operator to control multi-drone systems. We present a tool for generating tasks and plans for multiple drones in wilderness SAR scenarios. The central aspect of our approach is to improve the search quality by automatically generating tasks to ensure timely coverage of high-risk areas, such as ditches, lake and sea banks, and beneath tree lines, where distressed people are likely to be found.

Abstract Drones currently serve as a valuable tool for in-situ sampling of volcanic plumes, but they still involve manual piloting. In this paper, we enable autonomous dual plume sampling by using a machine vision model to detect eruptions. When an eruption is detected, a sampling trajectory is automatically generated to intercept the plume twice to collect comparative samples. The machine vision model is developed by training a YOLOv8 object detection model on a database of 1505 images featuring labelled plumes. The obtained average precision value of the model's plume class, at 90.7%, is comparable to that of state-of-the-art models for wildfire smoke monitoring. The performance of this method is assessed using a software-in-the-loop simulation of the drone and a simulated plume model. Although the results confirm the efficacy of using a machine vision model for triggering an onboard path-planning algorithm, they also suggest the potential for a hybrid strategy that integrates visual servoing with our proposed path-planning approach.

Abstract One of the main challenges in the real-world adoption of multi-Uncrewed Aerial Vehicle (UAV) systems lies in the specification of operations and the management of dynamic tasks in varied operational contexts. In this paper, we propose a multi-UAV planning architecture to reduce the level of specialized expertise necessary for handling multi-UAV systems. Furthermore, this work is the first step towards designing a multi-UAV planning architecture that integrates with the U-space services specified in EU regulation 2021/664. We propose two declarative languages: an Agent-Language for expressing mitigation and safety objectives for individual UAVs, and an Operation-Language to enable users to plan high-level multi-UAV operations based on the available resources. The languages enable automatic on-the-fly re-planning if any UAVs abort the mission unexpectedly. The initial result of the multi-UAV planning architecture is showcased in three simulated UAVs running as Software-In-The-Loop (SITL).

Abstract Multi-UAV systems have significant potential to enhance search and rescue (SAR) operations, since a search area can be covered faster than current approaches when multiple UAVs operate in parallel. While recent advancements within the field of multi-robot coverage planning have yielded promising results, current algorithms are predominantly centralized. In this paper, we present a generalization of the well-known decentralized consensus-based bundle algorithm (CBBA) that enables efficient task allocation in multi-UAV SAR operations. The generalized algorithm considers tasks as trajectories between two points where the traversal direction for each task is optimized in the task allocation process. We carry out a series of simulation-based experiments on benchmark problems and compare our results to a state-of-the-art centralized solution. We find that our novel decentralized approach yields times to completion similar to those achieved with a centralized coverage path planning approach, with only 1.9% overhead cost. We furthermore find that our approach performs 6% better than point allocations while scaling well with the number of UAVs involved in the search effort.

Abstract Emergency services organizations are committed to the challenging task of saving people in distress and minimizing harm across a wide range of events, including accidents, natural disasters, and search and rescue. Given the risks and the time pressure of these missions, adopting new technologies requires careful testing and preparation. Drones have become a valuable technology in recent years for emergency services teams employed to locate people across vast and difficult to traverse terrains. While an individual drone can be helpful by quickly offering a bird's eye view, future scenarios may allow multiple drones working together as a swarm to reduce the time required to locate a person. We conducted interviews as well as initial user studies with relevant stakeholders to understand the challenges and opportunities for drone swarms in the context of search and rescue. We distill our findings into five key research challenges: visualization, situational awareness, technical issues, team culture, and public perception.

Abstract Large-scale multi-robot systems have numerous potential real-world applications. It is, however, still unclear how a human operator can effectively engage and control a system composed of multiple autonomous robots, especially in unstructured and outdoor environments. This paper reports on ongoing work in the project HERD — Human-AI Collaboration: Engaging and Controlling Swarms of Robots and Drones — in which we focus on two concrete use cases from industrial partners, namely farming and search & rescue. One partner, Agro Intelligence ApS, sells autonomous farming robots; the other, Robotto ApS, develops autonomous drone-based monitoring solutions for emergency responders. Both partners aim to scale their technologies to multi-robot/multi-drone operations. In this paper, we present the two use cases, their differences and similarities, challenges and preliminary results.