Semi-closed industrial sites are a key emerging domain for applying connected and automated vehicle technology. The motivation for this includes increased productivity and continuity of operations. The safety challenges are more manageable than on public roads since only trained personnel with safety equipment are allowed to work at the site, and the driving speeds are often modest . This led to the development of an on-board software (SW) platform with safety features and the capability to operate in mixed traffic. The demo included two pilot cases: The Kouvola case and the Rauma case.

The Kouvola case (SC1-D01) focused on automating container loading and unloading to/from rail wagons with a reach stacker, complemented by autonomous container transport with an autonomous terminal tractor. Field tests with the Terminal Tractor and Reach Stacker were successfully conducted in Tampere, Finland, and in Ljungby, Sweden.

The Rauma case (SC1-D02) involved automating log transport with a sawmill area. The goal was to transfer log cages between various loading and unloading locations using an Automated Guided Vehicle (AGV). The AGV vehicle and the log cages to transfer are ready for fully computerized control.

Besides the vehicles, their remote operator support was also implemented. This support includes database response models, Prompt-to-SQL model training, and the remote operator portal including integration of Reach Stacker data playback and direct SQL-queries with a Large Language Model

response.

During the demos, two different sensor suits were tested: Lidar + camera during the automated driving, and lidar + spreader cameras during the spreader alignment and automatic container handling.

The safety aspects of the self-operating vehicles in both demo sites were analyzed by applying Preliminary Hazard Analysis (PHA) and System Theoretic Process Analysis (STPA). Also the approach described in UL4600 – Safety Standard for Autonomous Vehicles – was applied.

Kalmar Reach Stacker with autonomous loading and unloading of containers.

SC1: Safe Co-existence of Automated and Manual Transport at Industrial Sites

SC1.2 Rauma

Demo leader: Unikie

Partners involved: Mantsinen, Metsä Group, VTT

As part of the project, several different use cases were implemented. This case focuses on Metsä Group’s Rauma sawmill, where the goal was to explore autonomous log transportation in an industrial setting. The aim was to develop a solution that could automatically transport logs from the train unloading area to the sawmill’s feed table.

The solution is based on Unikie’s autonomous driving system, which uses sensors installed around the factory area to create a real-time digital twin of the environment. This system controls a ground vehicle developed by Mantsinen, capable of transporting log cages while monitoring surrounding activity. Operational safety was a key research focus due to the semi-open nature of the environment, where other vehicles and pedestrians may be present.

After three years of research and development, the concept was successfully tested in October 2025 at the Rauma sawmill by Unikie and Mantsinen. The site was equipped with sensors, cabling, and a server that enabled the ground vehicle to drive from behind the drying building to the sawmill’s feed table. A real-world use case was tested—albeit on a smaller scale—by transporting full log cages to the unloading area and returning empty cages to the loading point. Safety was also tested by walking in front of the vehicle, and driving precision was validated with sub-10 cm accuracy when positioning under the log cages.

All tests were successful, despite installation challenges due to a tight schedule. The ground vehicle was able to repeatedly position itself under the log cages and transport full loads along a predefined route. It stopped reliably—whether empty or loaded—when a person entered its safety zone. The concept worked as intended, and no obstacles were identified that would prevent commercial deployment.

We are proud to say that this is one of the world’s first solutions where a heavy ground vehicle can move autonomously and safely among people and manually operated machines in an industrial environment.

Testing envrionment at Metsä Group’s Rauma sawmill.

Testing operational safety with full log cage.

SC1.2

SC2: Search & Rescue (SAR) and Emergency Response for Civil Safety

SC2.1 Single-robot

Demo leader: Virtual Vehicle Research

Partners involved: AVL, HUAWEI Technologies Sweden AB, Pumacy, Universidad de Alcala, Montanuniversitaet Leoben, Ostbayerische Technische Hochschuleamberg-Weiden, AIT Austrian Institute of Technology, Bundesministerium fuer Landesverteidigung und Sport, EDI - Institute of Electronics and Computer Science, Laabmayr

Demonstrator SC2.1 represents a robotic platform capable of supporting Search and Rescue (SAR) operations in subterranean and GNSS-denied environments. As part of SC2, its purpose is to validate autonomous localization, mapping, and survivor detection capabilities under real-world underground conditions at the tunnel facilities at Zentrum am Berg (ZaB). The platform combines robust hardware, advanced perception technologies, and AI-driven algorithms to enable autonomous mission execution in challenging, GPS-inaccessible tunnel systems.

SC2.1 is an Unmanned Ground Vehicle (UGV) equipped with a suite of specialized sensors (e.g. LiDAR, thermal camera) and an onboard computing platform tailored to execute enhanced algorithms enabling autonomous navigation and AI-based survivor detection for complex underground missions.

Due to an extensive simulation study, robust open-source SLAM algorithms dealing well with featureless environments were selected to be implemented on the UGV. The Survivor Detection pipeline aims to detect and estimate the poses of potential human survivors near the robot platform.

Two approaches are introduced, an onboard method utilizing thermal images and LiDAR pointcloud, and an offboard method leveraging only thermal images. Data communication interfaces are established to transmit survivor positions and mission updates to emergency personnel. A wireless system simulator provides information about relevant communication attributes within the tunnels.

Additionally, SC2.1 supports a complementary feasibility study of magnetic field-based mapping and self-localization, evaluating the potential of magnetometer signals as an alternative positioning method in underground settings. This approach could, in the future, significantly enhance localization robustness where other signals are unreliable.

SC2.1 combines multiple innovative elements that push the state of the art in subterranean SAR robotics:

Autonomous operation in GNSS-denied subterranean environments: SC2.1 successfully performs localization and supports navigation without GNSS which is an essential requirement for tunnel missions.
AI-supported survivor detection: Integrated perception algorithms allow the UGV to identify survivors and provide their estimated positions to human responders, enabling faster and safer rescue operations.
Quantum sensing supported localization (future integration): In the context of SC2.1 the incorporation of a quantum sensor from SC5 will further elevate localization precision under highly degraded conditions.
Real-world validation: SC2.1 is deployed at the ZaB facilities, a unique European-scale experimental site offering realistic tunnel geometries and material properties.

These innovations allow SC2.1 to act as a reference platform for safe, autonomous, and efficient underground operations. The capabilities of SC2.1 are directly applicable to multiple real-world scenarios where human access is dangerous, slow, or impossible, such as SAR missions in collapsed e.g. tunnels, exploration and inspection tasks in GNSS-denied environments, safety assessment afterdisasters.

SAR real-world demonstrator vehicle at ZaB.

Visualization of the position of the SPIDER and the video stream from the UGV visualized in the XR COMMAND tool (provided by LAAB).

SC2.1

3.3

SC2: Search & Rescue (SAR) and Emergency Response for Civil Safety

SC2.2 Multi-robot

Demo leader: Virtual Vehicle Research

Partners involved: Universidad de Alcala, Montanuniversitaet Leoben, Ostbayerische Technische Hochschuleamberg-Weiden, AIT Austrian Institute of Technology, Bundesministerium fuer Landesverteidigung und Sport

Demonstrator SC2.2 entirely operates in simulation representing a simulation-based robotic platform and the operation environment within SC2 focusing on cooperative multi-robot system. A high-fidelity simulation model of ZaB and SC2.1 were developed enabling two or more UGVs to collaboratively localize and map a subterranean environment. The virtual attribute of the demo offers scalable and flexible development opportunities for testing advanced multi-agent strategies and communication-oriented behaviours without the logistical constraints and safety risks of physical deployment. The Demo comprises a complete simulation environment that models underground tunnel conditions, UGV operation, and communication effects. A detailed model of the ZaB tunnel facilities based on LiDAR measurements was developed within SC2.2, representing connected highway and railway tunnel tubes. Two UGV models are available in simulation, a Turtlebot4 platform and the SPIDER vehicle, including designated sensor characteristics. The multi-robot localisation and mapping modules allow at least two agents to build a shared representation of the tunnel environment. The effects of communication on multi-robot SLAM systems are covered in a conducted simulation study focusing on communication delay, de-synchronization and packet loss. Mitigating communication link failures between robots and/or a communication centre, an AI-based Quality of Service (QoS) prediction model is developed, providing forecasts of communication performance and blind spots at different locations.

SC2.2 introduces several innovations that advance the state of the art in subterranean SAR robotics:

Cooperative multi-robot SLAM in GNSS-denied environments: SC2.2 demonstrates how multiple UGVs jointly localize and map the ZaB tunnels without GNSS.
Communication-aware decision-making: A future integration of the QoS prediction into the navigation and coordination modules of the UGV will allow the agents to adjust their strategy to maintain reliable data exchange.
Simulation framework for subsurface SAR scenarios: SC2.2 provides a flexible and scalable validation and testing environment for evaluating algorithms and models before deploying them onto real systems, reducing development risks and costs.

The SC2.2 simulation-based demonstrator offers broad potential for applicability reaching from SAR research and robotics development over to safety-critical planning, such as research on GNSS-denied localisation and sensor fusion; pre-mission planning and optimisation before first responders enter hazardous environments; training and validation of AI-driven multi-agent coordination algorithms; etc.

SPIDER operating in one highway tunnel tube executing a SLAM algorithm (Fast-LIO2).

Example of reconstructed map using multi-robot SLAM with a fleet of 4 robots.

SC3: Digital Health and Emergency recognition for Driver and Operator

SC3.1 Detect and Save (DAS) System

Demo leader: AVL TR

Partners involved: AVL, AVL TR, EMO3D, ITML, SAT, POLITO

Demonstrator SC3.1 focuses on the Detect and Save system, which connects driver state monitoring with automated vehicle functions. The demonstrator is based on a Level 2 autonomous vehicle platform equipped with perception, lateral control and longitudinal control capabilities. Its purpose is to detect critical driver states such as fatigue, drowsiness or distraction, and to support a safe vehicle reaction when the driver is no longer able to operate the vehicle.

The system uses multiple sensing sources to monitor the driver and the vehicle environment. Smartwatch-based physiological data provided by SAT and POLITO is used together with in-cabin camera information from EMO3D’s CabinEye system. AVL contributes dashboard monitoring and surroundings perception components, while ITML provides the data fusion, flow and management infrastructure for transferring, storing and visualising relevant driver and vehicle data. AVL TR integrates these inputs with the demo vehicle and the automated driving functions.

When a critical driver state is detected, the Detect and Save system can activate the automated driving mode of the vehicle. The vehicle evaluates the surrounding environment and performs a controlled response, including lane keeping, movement towards the emergency lane or right lane, and a safe stop. In this way, the demonstrator goes beyond warning-based driver monitoring and links digital health assessment directly with vehicle-level safety functions.

The system was first evaluated in a CARLA-based highway simulation environment, where driver-state inputs were connected to the vehicle control logic. In the tested scenarios, fatigue or drowsiness-related signals triggered the Detect and Save function and the vehicle executed the intended emergency maneuver.

In addition to simulation, vehicle-level tests were also carried out in a controlled area. During these tests, the in-cabin camera setup and smartwatch-based driver status inputs were integrated with the vehicle computer. Sleep or distraction states were detected by the system and, after checking environmental safety, the vehicle control software guided the vehicle and brought it to a safe stop.

SC3.1 demonstrates an integrated approach for digital health and emergency recognition in automated vehicles. It combines physiological monitoring, camera-based driver observation, dashboard monitoring, surroundings perception, vehicle control and cloud-based data handling in a single demonstrator setup. The concept can be applied in future driver monitoring systems, automated driving safety functions, fleet monitoring solutions and emergency assistance concepts. It is relevant for OEMs, Tier-1 suppliers and research partners working on digital health, driver safety and automated emergency maneuvers.

Driving Test With Cabin Interior camera sensor in the Demo Vehicle.

End-To-End Integration Test Within Carla Simulation Environment.

In-cabin sensing driver's physiological and drowsiness signals visualized in Grafana dashboard.

SC3: Digital Health and Emergency recognition for Driver and Operator

SC3.3 Physiological and Air Quality Sensing

Demo leader: NVISION Systems and Technologies (NVISION)

Partners involved: NVISION, SAT, POLITO

The SC3.3 demonstrator shows how physiological and in-cabin air quality sensing can be combined to support safer and more human-centred monitoring. The demonstrator was developed as an integrated digital health system for real-world driving scenarios, where the driver’s condition is assessed together with environmental exposure inside the vehicle cabin.

The system brings together two complementary sensing streams. SAT and POLITO provide a Monitoring System based on commercial wearable technology (GARMIN smartwatch), able to acquire physiological variables such as heart rate, heart rate variability and other derived driver-status indicators. NVISION integrates the MICA WELL InBiot air quality sensor to continuously monitor relevant cabin parameters, including CO2, particulate matter, volatile organic compounds, temperature and relative humidity. The data streams are exchanged using common interfaces, including MQTT and JSON-based messages, enabling continuous communication between partner modules.

The innovation of the demonstrator is the combination of driver physiology and environmental context in a single monitoring pipeline. Instead of treating drowsiness, stress or fatigue as isolated physiological events, SC3.3 explores whether cabin conditions may contribute to changes in driver status. The collected multimodal data are processed through correlation-based and time-shifted analysis methods to identify immediate and delayed relationships between air quality variables and driver stress or drowsiness indicators. This supports the development of predictive, edge-ready AI models that can anticipate unsafe driver states before they become critical.

Integration and testing were carried out through joint sessions between NVISION, SAT and POLITO. Software integration was first verified in Barcelona, followed by real driving tests in Turin under urban and rural conditions. Additional in-vivo and in-lab driving sessions in Barcelona helped extend the dataset and strengthen the validation of the end-to-end sensing workflow. These activities confirmed that heterogeneous physiological and environmental signals can be collected, synchronized, stored and analysed within a shared demonstrator pipeline.

The demonstrator could potentially be applied in advanced driver monitoring, connected vehicles, fleet safety, professional transport and future digital health applications in mobility. By linking exposure, physiology and driver alertness, SC3.3 provides a practical basis for improving in-cabin safety, supporting preventive interventions and enabling more personalized monitoring of driver well-being.

In-vehicle SC3.3 setup during real driving data collection, combining MICA WELL InBiot environmental monitoring with smartwatch GARMIN wearable-based physiological monitoring.

Example outputs of the SC3.3 analytical pipeline: correlation heatmap and time-shifted correlation analysis between cabin environmental variables and driver stress levels.

SC4: Propulsion health and availability in safety critical situations

SC4.2 Concept of distributed diagnostics of powertrain

Demo leader: BUT

Demonstrator SC4.2 presents a Distributed Diagnostic System (DDS) for fault detection in PMSM drivetrains. The motivation behind its development is the fact that modern drive systems continuously generate large volumes of operational data, which it is desirable to systematically store and make available for predictive diagnostics purposes. The system is therefore designed not only to evaluate measured data in real time, but also to archive it for subsequent analysis. The aim was not to build the hardware for the direct vehicle use but rather the system capable to test developed diagnostic algorithms located on different layers of DDS.

The architecture is divided into three layers — powertrain, edge (Raspberry Pi 5 with a 5G modem), and server (Kafka cluster) — with the diagnostic algorithms compiled as a shared dynamic library deployable on both x86 and ARM platforms. Four independently switchable diagnostic subsystems are implemented: stator current analysis (frequency-domain harmonic analysis combined with three-phase current sum analysis), rotor position sensor fault detection from stray flux measurements, DC link voltage sensor fault detection using a fully connected neural network implementing a PMSM voltage model, and stator winding interturn short fault detection.

The neural network was trained on 27 experiments covering different speed ranges and both torque directions, with a maximum estimated stator voltage error not exceeding 4 V. Experimental validations on a real PMSM with stray flux sensing and fault emulation confirmed reliable detection of all four fault types under the respective operating conditions, with data processing on the Raspberry Pi 5.

The block diagram of the DDS is shown in Figure 1. The inverter communicates with the Raspberry Pi edge device over UDP. Data are then transmitted via a 5G network to a physical server hosting the Kafka cluster, which handles both data storage and processing.

Block diagram of DDS

For the testing, the experimental setup was constructed as shown in Figures 2 and 3. Figure 2 shows the customized motor with integrated Hall flux sensors and the capability to emulate several fault types, connected to a dynamometer that enables loading the motor across a range of defined operating conditions. Figure 3 shows the other side of the setup, with the inverter, Raspberry Pi 5 edge device, and 5G modem.

The setup of inverter powerstage, inverter controller, Rpi 5 edge device and 5G modem

Final validation focused on a realistic deployment scenario in which the edge device transmits measurement data in batches with pauses between them, mimicking actual field operation. Figure 3 shows the diagnostic output for an experiment with an emulated DC link voltage sensor fault. The four subplots show, from top to bottom: the stator phase currents, the emulated fault profile together with the cumulated fault indicator, the individual fault detection indicators, and internal diagnostic signals including the rotor angle residual and stator voltage residual. The fault was successfully detected and reported, with the cumulated indicator rising consistently after fault onset. Computational performance was evaluated on both a desktop PC (Intel i7-4790K) and a Raspberry Pi 5. Processing times for one message (2500 samples) ranged from 12.0 ms with no fault detection algorithms active to 16.7 ms with all algorithms enabled on the desktop, and from 11.6 ms to 14.6 ms on the Raspberry Pi 5.

Plotted diagnostic data in development mode

SC4: Propulsion health and availability in safety critical situations

SC4.4 Digital twin of the vehicle

Demo leader: HSO

Partners involved: MBAG, BUT, AVL

A modified electric Motor equipped with flux sensors (SC 4.1) provides additional operational data that offers deep insights into the motor’s internal physical processes. These data are valuable for a variety of applications, including the development of an AI-based digital twin (Fig. a).

In this approach, a black-box neural network is trained to reproduce the relationships between measured inputs and outputs as accurately as possible. Given a sufficiently precise data basis, a highly accurate dynamic black-box model can be created that is capable of continuously improving its performance through further learning during operation.

However, accurate data alone are not sufficient to create a reliable digital twin. The training process must also be properly configured, and the raw measurement data require appropriate preprocessing. Within SC 4.4, the focus was therefore placed on developing a valid digital twin concept by investigating how different hyperparameters influence the training performance.
A Neural State Space (NSS) architecture (Fig. c) was selected as the AI framework, as it proved particularly effective in capturing the dynamics of time-series data.

The SC 4.4 demonstrator consists of a complete training and evaluation framework for NSS-based digital twins. It can be supplied with arbitrary datasets, either generated through simulation or measured on a test-bench (Fig. b), while allowing all relevant hyperparameters to be freely configured. In addition, the demonstrator includes automated parameter variation and subsequent performance evaluation.

The research revealed that, for the chosen architecture, network size plays only a minor role. Even relatively small networks with only a limited number of interconnected artificial neurons are capable of achieving high performance. Once a minimum network size threshold is exceeded, further increasing the number of neurons yields only marginal improvements. In contrast, the segmentation of training data into batches was found to have a significant impact on model quality. Both excessively long and excessively short batches lead to a noticeable degradation in performance. Furthermore, the results demonstrate that overlapping batches substantially improve prediction accuracy. The stronger the overlap, the more accurate the resulting digital twin becomes, including its ability to generalize to previously unseen datasets (Fig. d).

The findings of SC 4.4 provide practical guidelines for the efficient creation of digital twins for electrical motors. The approach enables highly accurate models to be developed with comparatively little effort and without requiring detailed knowledge of the motor’s internal structure.

Schematic representation of the workflow and development process.

SC2.2

4.4

3.1

4.2

SC4: Propulsion health and availability in safety critical situations

4.5

SCD4.5 e-Machine propulsion health and availability in safety critical situations

Demo leader: I&M

Partners involved: I&M, UNIMORE, TEKNE

The aim was to develop a demonstrator for the flux-based control system on a commercial permanent magnet (PM) concentrated-winding motor to improve controllability, torque delivery, and provide an intrinsic redundancy in the sensing elements installed into an e-machine (Figure 1):

Flux sensors installed on one side of the lamination stack on dedicated PCBs, one for each stator tooth.
Automotive-grade electronics used both on the Hall-effect sensors and in the motor drive.
Inverter built on a modular inverter architecture developed by one of the partners, equipped with extra differential reading channels to acquire the signals of the flux sensors installed on the motor, limiting the effects of the EM noise.

The demonstrator was installed on a testbench to characterize the flux sensor signal across a wide range of operating points and considering three different sensor positions: one in the middle of the stator tooth, one near the airgap, and one facing the rotor. Each sensing position refers to a three-phase triplet of signals for the real-time detection of the phase angle of the measured quantity.

Innovation

The comprehensive mapping of flux signals across different sensor positions and operating points (Figure 2) provided a complete overview of the type and quality of the information retrievable in concentrated-winding PM motors. From this analysis, rotor position and PM flux estimation from the rotor-facing sensors emerged as the most promising use of the flux information. This led to the development of a novel, flux-based observer for the estimation of these quantities, allowing for the potential replacement of more expensive position sensors while providing additional information on the PM temperature and/or magnetization state, which can be used for health-monitoring purposes. In addition, the PM flux estimation enables the use of temperature-corrected models for more accurate torque control both with traditional Field Oriented Control (FOC) and Direct Flux Vector Control (DFVC) algorithms.

Application

The advantages of this solution are multiple and can be applied both in the automotive and in the aerospace domains:

The presence of flux sensors can eliminate the need for an external position sensor, reducing the rotor imbalance and enabling more compact implementations.
The magnet temperature estimation makes it possible to implement health-monitoring and prognostic algorithms able to improve the life expectancy of the powertrain and intervene in case of unexpected degradation.
Multiple sensors can be installed in a redundant configuration to improve reliability in safety-critical applications.

The components used into the demonstrator and their final combination into the testbench

Mapping of flux sensor signals over various Id-Iq operating points (triplet of rotor-facing sensors)

SC5: Quantum Sensor Multi-Modal, Multi-Physical Sensing at Highest Precision

SC5.1 Quantum sensing using NV-Centres in diamond

Demo leader: Institute of Electronics and Computer Science (EDI)

Partners involved: AVL (FAU Erlangen-Nürnberg), ARQUIMEA Research Center

Quantum technologies are commonly grouped into three major domains: quantum computing, quantum communication, and quantum sensing. While quantum computing promises transformative capabilities for solving complex computational problems and quantum communication aims to provide fundamentally secure information transfer, quantum sensing is currently the most mature of the quantum technology domains and the closest to broad practical adoption. A-IQ Ready brings quantum sensing from laboratory prototypes to industrial and commercial applications. Specifically, we focus on Nitrogen Vacancy (NV) centre defects in diamond, which can be used to develop ultra-precise magnetometer sensors, and have the unique capability to operate at room temperatures. We join our interdisciplinary expertise to:

improve “out-of-the-box” usability of the sensor,
reduce its size and costs,
and achieve the world's fastest readout (while preserving advantageous performance).

A-IQ Ready’s quantum sensor system and sensor miniaturised fibre-based sensor head design.

Measurements for underground/indoor localisation

In our recent indoor localisation tests, we explored the potential of quantum sensor magnetometers for precise spatial measurements. The study was conducted over two iterations to assess the reliability and accuracy of the technology in dynamic environments. For the second iteration, we took a step further by automating the process, leveraging a ground-truth OptiTrack positioning system alongside a robot to ensure consistent and repeatable measurements. This enhanced approach allowed us to refine our data collection methods and minimise human error. In the data, you can observe a magnetic field anomaly, likely caused by a structural element of the building, which highlights the sensitivity of the quantum sensor in detecting even the smallest variations in the environment. This anomaly is an interesting finding, demonstrating the complexities involved in indoor localisation and the power of quantum sensors in uncovering hidden details within physical spaces.

Setups for initial and automated magnetic field mapping tests. (a) Initial tests at the premises of VIF; (b) Sensor deployment on automated robot platform; (c) Indoors magnetic field mapping at EDI’s automated testing rig.

Measurements for motor control

This motor control use case presents a fascinating challenge, as it involves a quantum sensor that requires an exceptionally high framerate for accurate measurements. To tackle this, we integrated the fibre-based sensor head of the quantum sensor directly inside a motor, a complex task that required precise engineering. During our tests, we gathered data that demonstrated the sensor's ability to track and determine the motor's phase. This breakthrough is significant, as it opens up new possibilities for using quantum sensors in real-time motor diagnostics and control systems.

Measurement setup for motor control tests using a quantum magnetometer. (a) Measurement setup; (b) Measurement data illustrating the motor’s rotation phase; (c) Fibre-based sensor head concept.

Next steps

The advancements achieved in our quantum sensor development already provide a highly valuable input for the continued evolution of the technology, particularly within Chips-JU Horizon Europe initiatives such as ARCHIMEDES, Cynergy4MIE, and MOSAIC. Insights gained from these projects have directly informed the refinement of both system design and implementation strategies. Notably, the analogue electronics circuitry we developed is now undergoing miniaturisation, marking an important step toward more compact and scalable solutions. In parallel, we have introduced an optimised event-driven control hardware architecture, designed to enhance efficiency and responsiveness. Together, these developments represent a significant stride toward practical, deployable quantum sensing systems that can meet the demands of real-world applications.

Miniaturisation effort of quantum sensor’s analogue electronics.

SC5.1

SC7: Cooperative Multi-Agent Systems (Decentralized AI for Emergent Industrial Solutions)

SC7.1 AGV system

Demo leader: Rosenheim University of Applied Sciences

Partners involved: Safelog, ScaliRo

The technical system of Demonstrator 7.1 has been successfully implemented in the laboratory environment of Proto_lab. In close collaboration between Safelog and THRO, the automated guided vehicles (AGVs) and the necessary peripheral components, such as charging stations and actuators for automated gate control, have been installed and tested. This step represents a significant advancement in the development of the AGV system. As part of the implementation, a shift from WIFI to 5G communication infrastructure was carried out. This measure aims to substantially improve the performance of data transmission between various hardware components and the control system. Tests have demonstrated that 5G technology offers stable connections and faster data rates, which are crucial for the efficiency of the overall system.

The cloud-based control system was established by our project partner ScaliRo. In this context, THRO programmed the travel paths, routes, and internal processes to compare them with the reference use case FT06. Continuous testing during the implementation phase allowed for further optimization of the overall system performance.

A standout result of this phase is the reduction of the throughput time using a standard controller to manage a FT06 pass. With a seamless operation, the throughput time was optimized to approximately 1 hour and 50 minutes. This represents a significant improvement in efficiency. Additionally, the use of the controller from WP 4, which includes a newly developed AI controller, was also successfully tested in the simulation environment of the ScaliRo Cloud. The performance of this new technology was slightly better in terms of throughput time compared to the previously used standard controllers from ScaliRo.

Overall, the current status of Demonstrator 7.1 shows that all technical components are working seamlessly together, and the set goals regarding system optimization have largely been achieved. The close collaboration of project partners has proven vital for the success of this implementation. The next steps will focus on integrating the insights gained into the further development of the AGV system and conducting additional tests to validate stability and performance under real-world conditions.

Setup of the training environment.

Technical implemented fleet of AGVs in the laboratory environment.

SC7.1

SC7: Cooperative Multi-Agent Systems (Decentralized AI for Emergent Industrial Solutions)

SC7.2 Virtual Training Platform for AMR Navigation in Unstructured Warehouses

Demo leader: Technical University of Munich

Partners involved: AIT Austrian Institute of technology, Technical University of Graz

Overview:

We deliver a modular virtual training platform that trains, evaluates, and verifies reinforcement-learning (RL) navigation policies for autonomous mobile robots in challenging warehouse environments. The toolchain integrates a Unity-based digital twin, Nav2-inspired model-based controllers, specification-driven reward shaping, and ontology-based scenario generation into a reproducible pipeline for offline-to-online RL and

automated validation.

Developments:

Simulation & tooling (TUM): Procedural Unity environments with configurable physics, lighting, and appearance randomization; multi-modal sensors (RGB BEV/front, LiDAR, IMU, experimental quantum magnetometer model); deterministic annotations and a gRPC API for state/action exchange. Packaged as Docker images for on-premises use.
Control & data pipeline (TUM): ROS2-decoupled Nav2-inspired controllers (A*/local planner/DWA) for baseline behavior and supervised trajectory logging to bootstrap offline RL (TD3-BC) and subsequent online fine-tuning (PPO/SAC/online-TD3).
Specification & monitoring (AIT): Hierarchical potential-based reward shaping (HPRS) and hybrid runtime monitoring (approximate + exact) that translate formal temporal requirements into dense rewards and automated oracles for safety checks.
Ontology & test generation (TUG): Browser app to author ontologies, convert constraints into parameter sets, combinatorial generation (pairwise/3-wise), OpenSCENARIO bundles for compact, reproducible test suites and curriculum progression.

Key innovations:

Structured realism: semantically constrained domain randomization (color palette spacing, probability-weighted lighting, physically coherent parameter ranges) to produce realistic, diverse training distributions that improve transfer.
Disturbance & delay awareness: explicit disturbance channels and delay hooks in simulation, plus disturbance-aware representations in policies for more robust, delay-resolved control.
Spec-to-reward integration: automated pipeline from formal requirements to prioritized shaped rewards, reducing manual reward engineering and enabling verified training.
Reproducible, semantic testing: ontology-driven scenario generation yields machine-checkable regression suites and critical-scenario libraries for safety validation.

Applications:

Intralogistics: AGV/AMR navigation, routing, docking, congestion management, and multi-robot coordination.
Research & education: reproducible RL benchmarks, lab courses, and student projects.
Industrial pre-validation: scenario-based risk assessment and pre-deployment testing for system integrators and operators.

Partner contributions for SC7 demo. 2.

Scalable simulations with canonical colour randomization for robust agent training.

Integration flow of components developed by each partner as a sequence diagram.

SC7.2

SC1: Safe Co-existence of Automated and Manual Transport at Industrial Sites

Go to SC1.1 Kouvola

SC2: Search & Rescue (SAR) and Emergency Response for Civil Safety

SC3: Digital Health and Emergency recognition for Driver and Operator

SC4: Propulsion Health and Availability in Safety Critical Situations

SC5: Quantum Sensor Multi-Modal, Multi-Physical Sensing at Highest Precision

SC7: Cooperative Multi-Agent Systems (Decentralized AI for Emergent Industrial Solutions)

SC1.1 Kouvola

SC1.2 Rauma

SC2.1 Single-robot

​

SC2.2 Multi-robot

​

SC3: Digital Health and Emergency recognition for Driver and Operator

SC3.1 Detect and Save (DAS) System

​

SC3: Digital Health and Emergency recognition for Driver and Operator

SC3.3 Physiological and Air Quality Sensing

​

SC4: Propulsion health and availability in safety critical situations

SC4.2 Concept of distributed diagnostics of powertrain

​

SC4: Propulsion health and availability in safety critical situations

SC4.4 Digital twin of the vehicle

​

SC4: Propulsion health and availability in safety critical situations

SCD4.5 e-Machine propulsion health and availability in safety critical situations

​

SC5.1 Quantum sensing using NV-Centres in diamond ​Demo leader: Institute of Electronics and Computer Science (EDI)

SC7.1 AGV system

​

SC7.2 Virtual Training Platform for AMR Navigation in Unstructured Warehouses ​​

SC5.1 Quantum sensing using NV-Centres in diamond

Demo leader: Institute of Electronics and Computer Science (EDI)

SC7.2 Virtual Training Platform for AMR Navigation in Unstructured Warehouses