Lessons Learned from the RAICAM Doctoral Network Research Sprints

Lessons Learned from the RAICAM Doctoral Network Research Sprints Alperen Kenan1, Sahar Sadeghi Kordkheili2, Juan Jose Garcia Cardenas3, Alessandro Melone4, Changda Tian5, Haichuan Li6, Hamidreza Raei7,11, Sasanka Kuruppu Arachchige8, Yifeng Tang9, Adriana Tapus3, Anibal Ollero2, Arash Ajudani7, Begoña C. Arrue2, Dimitrios Papageorgiou5,10, Joni-Kristian Kämäräinen8, Jukka Heikkonen6, Luis Figueredo4,12, Manuel Giuliani13, Panos Trahanias5,14, Paul Bremner1, Saeed Rafee Nekoo2, Simon Watson9, Tomi Westerlund6 1Bristol Robotics Laboratory, University of the West of England, UK [email protected] ⋆ Abstract. Doctoral Networks (DNs) aim to address systemic challenges in doctoral education, such as fostering interdisciplinarity, enabling international and intersectoral collaboration, enhancing employability, and promoting responsible innovation. While cohort-based training helps mitigate student isolation through workshops and summer schools, traditional DNs often struggle to fully realise their collaborative potential, often relying on predefined supervisor relationships or the initiative of individual researchers. In contrast, Marie Skłodowska-Curie Doctoral Networks (MSCA-DNs) prioritise doctoral candidates (DCs), challenging them to balance independent research with contributions to a shared, missiondriven objective. This study examines how structured training, including digital communities and application-focused research sprints, enhances system integration and collaboration within the Robotics and AI for Critical Asset Monitoring (RAICAM) Doctoral Network. DCs located across seven European countries worked in virtual teams, refining systems through structured workflows, weekly meetings, and shared workspaces before training schools. Through continuous online collaboration and targeted sprints, RAICAM facilitated interdisciplinary integration. Two research sprints, conducted in Italy and France, allowed teams to develop ⋆2 GRVC Robotics Lab, Universidad de Sevilla, Spain, 3 Autonomous Systems and Robotics Lab, Institute Polytechnique de Paris, France, 4 Munich Institute of Robotics and Machine Intelligence, Technische Universität München, Germany, 5 Institute of Computer Science, Foundation for Research and Technology Hellas, Greece, 6 TIERS lab, University of Turku, Finland, 7 HRI 2 Lab, Istituto Italiano di Tecnologia, Italy, 8 Computer Vision Group, Tampere University, Finland, 9 Department of Electrical and Electronic Engineering, University of Manchester, UK, 10 Department of Electrical & Computer Engineering, Hellenic Mediterranean University, Greece, 11 Department of Electronics, Information and Bioengineering, Politecnico di Milano, Italy, 12 School of Computer Science, University of Nottingham, UK, 13 Faculty of Electrical Engineering, Kempten University of Applied Sciences, Germany, 14 Computer Science Department, University of Crete, Greece. 2 A. Kenan et al. and test solutions for real-world challenges with an impact-driven plan that considers a given problem from and end-to-end perspective that requires and foster interdisciplinary collaboration. The results highlight the effectiveness of structured training in enhancing collaboration and adaptability, while identifying key areas for improvement. This study translates lessons from RAICAM into practical guidelines for future doctoral networks, demonstrating how structured training empowers students to drive interdisciplinary research independently. Keywords: Doctoral Network ·Research Sprints ·Multi-Robot Systems ·Interdisciplinary Collaboration ·Autonomous systems 1 Introduction The core principle underlying doctoral networks is to provide candidates with highquality research training in a collaborative, structured cohort environment. This model enhances career prospects in both academic and non-academic fields [ 1 ]. By promoting international, interdisciplinary, and intersectoral collaboration, DNs mitigate student isolation and facilitate knowledge exchange and secondments across diverse research settings. They also emphasize ethical research through Open Science principles and peer review, ultimately cultivating creative, resilient researchers capable of addressing real-world challenges [1]. The cohort-based training format emerged in the early 2010s with the establishment of Centres for Doctoral Training (CDTs) in the UK [ 9 ] and the MSCA Doctoral Networks in Europe [ 2 ]. Since 2014, the MSCA DN scheme has funded 1,475 programmes with a budget exceeding €400 million per year; in the UK, the 2023 scheme supported 65 new networks with a £1 billion investment. Practical DN objectives require close collaboration among PIs. In EU and international cohorts, the focus may shift from DC training to PI-driven projects, potentially isolating research efforts. Here, we present an alternative experience from the Horizon MSCA DN RAICAM project 1 , which adopts an impact-driven, application-focused research plan that considers the problem end-to-end. This approach fosters collaboration and facilitates the development of new tools such as the Digital Community and Research Sprints to achieve DN objectives. 1.1 RAICAM The Robotics and AI for Critical Asset Monitoring (RAICAM) DN was funded through the 2021 MSCA-DN scheme, formally starting in January 2023. The DN has 10 academic partners, each hosting a Doctoral Candidate (DC), across 7 countries as detailed in the author list. In addition, there are 4 industrial partners; FIS360 and Sellafield Ltd from the UK, the Fraunhofer Institute in Germany and Anybotics, based in Switzerland. The aim of RAICAM is to train the next generation of robotic systems engineers who will develop creative and innovative multi-disciplinary skills, with 1 Further details about the Horizon MSCA DN RAICAM project can be found on the official website: https://raicam.eu/ Lessons Learned from the RAICAM Doctoral Network Research Sprints 3 the scope of research focused on how autonomous robots could undertake sample retrieval missions (particulates or small debris) in industrial facilities [5]. Fig. 1: Pipeline of the training structure 1.2 Training Structure While doctoral networks generally follow a well-structured group-based learning model, RAICAM adopted an innovative approach to better realize the collaborative potential of this framework. Many cohort-based training programs encourage collective learning through workshops or summer schools, but interaction beyond these prescribed activities is often optional or unstructured. While collaboration is encouraged, it frequently depends on pre-existing supervisor relationships or the initiative of proactive students. Collaboration is most effective through an impact-driven, application-focused research plan that considers the problem from an end-to-end perspective. This approach was pioneered through events like the DARPA Robotics Challenges [ 6 ], which focused on exploring environments with fleets of heterogeneous robots. Successful teams addressed the problem as a whole, rather than as a series of individual technological challenges, leading to integrated solutions. To address these limitations, RAICAM DN implemented a unique structure centered on three core elements: (1) an overarching research challenge, (2) a digital community platform, and (3) annual research sprints. The pipeline of the training structure is shown in Figure 1. Overarching Research Challenge At the inception stage, it was decided that all PhD projects would have to contribute towards an overarching challenge, specifically autonomous robotic sample retrieval in industrial facilities. This challenge, co-defined by industrial partners, requires a multi-disciplinary approach to address. The PhD projects were split into three sub-categories as shown in Fig. 2: Environmental Interaction, Perception and Cognition, and Human-Robot Interaction. By co-creating complementary projects that aligned with a shared research goal, DCs significantly enhanced their collaboration. The supervisor team essentially mapped out potential collaborations in advance, working them into the required MSCA secondment plan, significantly enhancing the student’s opportunities for additional publications. 4 A. Kenan et al. Fig. 2: RAICAM research topics Digital Community and Annual Research Sprint Having an overarching research challenge does not guarantee collaboration, especially if students only meet sporadically for training workshops, symposiums or secondments. The most productive collaborations come from good personal relationships, which can be difficult to form if you are distributed across seven countries in Europe. To address this issue, two activities were implemented. The first was a regular weekly online meeting for all students. These informal, community-building sessions help students connect socially and technically. Weekly activities include wellbeing updates, research updates, mentoring, professional development, guest lectures and training, all undertaken in a relaxed and collegiate environment. Often in DNs, relationship building is left to the in-person activities, but in RAICAM, this is done before the students met in person. The second element was to give the students a focal point for collaboration. The overarching research challenge meant that each student was already working towards a common goal, but this was made more tangible through the concept of cohort-level demonstrations through annual research sprints. The concept of a research sprint evolved from hackathons, where students come together in-person for a time-limited activity (often 24 - 72 hours) where they solve an unknown challenge [ 3 ]. A key aim is to get people out of their day job environments to promote innovation. Hackathons promote interdisciplinary collaboration during the event, but the mediumand long-term benefits are often limited. This research sprint approach was adopted within RAICAM as the focal point to bring the cohort together to form a strong research community. The students are required to demonstrate progressively more complex demonstrations on an annual basis which showcases either integrated aspects of their research, or the skills they have developed. The pedagogical aim is to help the students develop project management, systems integration and field deployment experience. Combining the research sprints with the digital community significantly enhances the students’ learning experience, employability, and the scope for meaningful collaborations. Lessons Learned from the RAICAM Doctoral Network Research Sprints 5 2 Heterogeneous Multi-Robot Search and Intervention Missions This section outlines the overarching challenge addressed by RAICAM, guiding an impact-driven, application-focused approach, along with key sub-challenges. This mission-driven framework enables DCs to develop core robotics skills and integrate their research in an intersectoral, interdisciplinary context. The primary objective is to design a heterogeneous multi-robot search and intervention system for various missions, including the European Robotics Hackathon (EnRicH). The final demonstration is scheduled for 2026, at the conclusion of the RAICAM project. This paper presents the first two demonstrations, held at Istituto Italiano di Tecnologia (IIT), Italy (April 2024), and École Nationale Supérieure de Techniques Avancées (ENSTA), Paris, France (November 2024). Each demonstration lasted four days, with one DC attending the IIT event virtually. In preparation, DCs collaborated remotely for 3–4 months to develop and integrate hardware and software components. For each research sprint, a project plan and workflow were developed, with the mission subdivided into work packages with allocated resources. The plan included sub-sprints, on-site workshops, and post-workshop phases. Tasks covered technical areas such as simulation, mapping, navigation, and manipulation, alongside non-technical aspects like logistics, presentations, and publishing. Coordination was managed through a Notion 2 workspace, which also served as a platform to organise the demonstrations. Within this space, the DCs created dedicated pages to coordinate their demonstrations at IIT and ENSTA 3 . They were tasked with designing the demonstrations and given freedom to explore any concept within the multi-robot collaboration framework. The researchers consulted with their supervisors and provided regular updates during community meetings, where academics offered guidance. Guest lecturers delivered training on project management and field robotics to help address major challenges. 2.1 Technical System Overview The initial mission for the first demonstration at IIT involved the autonomous detection, localisation, and interaction with a battery, including episodic voltage measurement by a robot. This was extended for the second demonstration at ENSTA Paris, where the final task involved manipulating industrial valves. Fig. 3 shows an overview of the system, which comprises an aerial drone (Agipix [ 7 ], with a Livox Mid360 3D LiDAR), a legged robot (Unitree Go2 [ 11 ]), and a ground robot equipped with a robotic manipulator (Clearpath Husky with UR5 manipulator [ 10 ], Robotiq 2F-140 gripper, and Intel D435I RGB-D camera). ROS2 was used as the middleware for the system, although several components still used ROS1, necessitating the use of RosBridge. The Foxglove web-based visualisation platform was used to monitor critical telemetry of the robots. 2Further details about the Notion workspace can be found at: www.notion.com 3 The Notion workspace pages for the IIT and ENSTA demonstrations can be accessed via: IIT Demonstration and ENSTA Demonstration. 6 A. Kenan et al. Fig. 3: Overview of multi-robot collaboration: The Agipix drone maps and transmits a cloud map to the Unitree Go2, which locates a valve and sends its position to the Husky-UR5, which then detects, estimates, and manipulates the object. 2.2 Simulation Structure and Digital Twin: Core Enablers for Collaboration To enable remote collaborative development, testing, and validation of multi-robot systems, a unified simulation platform integrating all necessary components was required. A shared virtual environment allows remote teams to work synchronously on various aspects, such as perception, control, and coordination strategies, without hardware constraints. Access to the same environment and robot models is crucial for all DCs to ensure consistency between teams. NVIDIA Isaac Sim [ 12 ], built on the Omniverse platform, was chosen for its high-fidelity physics and photorealistic rendering. Cloud-based systems and shared code enable real collaboration, fostering knowledge exchange. To create a comprehensive multi-robot simulation environment, several platforms were integrated, as shown in Fig. 4. This modular framework and standardised interfaces ensured ease of use, enabling cross-border teamwork. Drone Simulation The Pegasus Simulator[ 4 ], a high-fidelity aerial robotics framework integrated with Isaac Sim, was used for drone simulation. It provides a custom Python control interface, enabling drone control within IsaacSim using the PX4 control stack. IsaacSim’s sensor framework also enabled the integration of the LiDAR. Figure 4 (a) and (b) show the relevant simulation details. Legged Robot Simulation The Unitree Omniverse was used to model the GO2 quadruped in Isaac Sim, featuring locomotion control on various terrains, force and torque simulation with PhysX, and LiDAR and depth cameras for terrain perception (Figure 4 (c)). Mobile Robot and Manipulator Simulation Two ground-based configurations were implemented: the Summit XL mobile robot with a Panda arm and the Husky robot with a UR5 arm. Both combine mobile control with dexterous manipulation (Figures 4 (d, e)). The platforms are fully integrated with ROS2 control frameworks, enabling seamless communication between controllers and sensor data processing pipelines. Lessons Learned from the RAICAM Doctoral Network Research Sprints 7 (a) Drone modeling (b) Simulation for our drone. (c) Simulation of the legged robot Go2. (d) Simulation for IIT demo. (e) Simulation for Husky and UR5. (f) Creat sim-env from matterport Fig. 4: The simulation platform for multi-robot system utilized in the demonstration. Digital Twin Construction Matterport was used for 3D mapping and reconstruction to ensure fidelity between simulated and real environments. This approach captures real-world spatial structures and converts them into virtual environments in Isaac Sim, creating high-resolution digital twins for demonstration scenarios. Leveraging digital twin technology, the simulated environments provide an accurate testing platform before deployment (Figure 4 (f)). Object Detection Robust object detection is crucial for real-world autonomous navigation, especially under visual disturbances. To address this, a comprehensive dataset was created, combining real-world images with simulated data from Isaac Sim, capturing an object’s unique characteristics (a valve in this case). AI-assisted tools like SAM 2.0 (Segment Anything Model) were used for efficient annotation and fine-tuned with a pre-trained YOLO v5 model [13]. Mapping and Navigation The LiDAR point cloud and IMU frames are fused using FASTLIO2 [ 14 ] to generate the 3D map. Registration and loop-closure with Scan Context [ 8 ] produce the final map and high-speed real-time odometry for indoor drone localisation. At the survey’s end, the 3D map is shared over the network for use by other robotic platforms. The goal is to navigate an unknown environment while building a map for subsequent operations. FAR Planner [ 15 ] dynamically updates a visibility graph for real-time path re-planning. 3 Residential Research Sprint Demonstrations During the training school activities at both IIT and ENSTA, the goal was to perform missions replicating nuclear industry scenarios, combining various levels of shared autonomy. DCs aimed for multi-robot collaboration by integrating 8 A. Kenan et al. (a) Researchers and their supervisors collaborating at IIT. (b) Researchers with the hosting PI during collaboration at ENSTA. Fig. 5: Group photos from IIT and ENSTA demonstrations, showcasing researchers and their supervisors. their research topics, such as mapping, navigation, and manipulation, towards a common goal, without human intervention, as in real-life scenarios. With only 3–4 days for in-person system integration and demonstrations, the research sprints relied on RAICAM’s modular architecture (Section 2.1) and shared simulation infrastructure (Section 2.2). These frameworks enabled virtual collaboration, aligned with the end-to-end challenge, while also allowing localized hardware testing. This hybrid approach supported RAICAM’s goal of integrating heterogeneous robotic systems into a cohesive solution. For the DCs, key learning experiences focused on multi-robot integration and transitioning from simulation to real-world environments. Within the frameworks of Sections 2.1 and 2.2, they developed skills in aligning sub-challenges with the broader mission. Figure 5 shows the participants in the demonstrations. 3.1 Voltage Inspection Task: Demonstration at IIT In the first demonstration at IIT, the DCs performed an inspection task using shared autonomy across a fleet of robots to measure electric voltage. The scenario involved a UAV transported by a mobile robot, which mapped the environment and identified the area of interest. The mobile robot then autonomously navigated to the target location, avoiding obstacles. Upon arrival, an operator used a teleoperation interface with stereo cameras and an IMU for visual-inertial odometry to guide the robotic arm, which grasped the tool and positioned it on circuit points for voltage measurement. An impedance controller ensured stable interaction despite environmental uncertainties. The subtasks involved are shown in Fig. 6, including UAV deployment 6 (a), environmental mapping 6 (b), simulation of multi-robot scenarios in IsaacSim 6 (c), autonomous ground robot navigation 6 (d), teleoperation interface for tool grasping 6 (e), and voltage measurement 6 (f). Lessons Learned from the RAICAM Doctoral Network Research Sprints 9 (a) UAV and ROV initial position. (b) Mapping done by the UAV. (c) Simulation of the robots. (d) Ground robot navigation. (e) Manipulator teleoperation. (f) Inspection of the voltage. Fig. 6: The multi-robot system utilized in the demonstration at IIT. 3.2 Valve Manipulation Task: Demonstration at ENSTA Paris During the research sprint at ENSTA Paris, the DCs extended the IIT activities to demonstrate a valve manipulation task. The drone mapped the environment for situational awareness, the legged robot localized the target valve, and the UR5 manipulator on the Husky platform performed valve detection, pose estimation, and manipulation. These sub-tasks were developed and validated in simulation, enabling rapid transition to physical robots. The integration phase allowed the DCs to refine multi-robot coordination protocols, gaining insights into sensor measurements, navigation, hardware-software interoperability, and teamwork. This collective effort deepened their understanding of optimizing complex, crossplatform robotic solutions for industrial and field applications. The subtasks involved in the ENSTA scenario are illustrated in Fig. 7 (a) and (b), with the drone and legged robot initiating exploration. The primary objective is to survey and generate a 3D map with the drone, used for task identification by other robots. In the next step, as shown in subplots 7 (c) and (d), the target area is identified using LiDAR odometry and a real-time RGB camera feed. The drone continues surveying until a complete map is generated and the target is identified, then communicates the location to the fleet via ROS2. Meanwhile, the legged robot optimizes exploration, easing the Husky’s task of locating the area of interest. For object detection, the YOLOv5 model, trained with SAM2-annotated datasets, detects the valve, as shown in subplot 7(e). Finally, in subplot 7(f), the valve operation is executed by the Husky. 3.3 Research Spring Summary and Challenges In the first scenario, the primary task was voltage inspection, involving a drone and a ground robot with a manipulator. Task evaluation in the unified simula-