The experiment addresses critical operational and technical challenges in the current maintenance management landscape, specifically the “disconnect” between field operations and digital systems. Currently, maintenance personnel rely on Human-Machine Interfaces (HMIs) that display high volumes of notifications, which are often ignored due to information overload. There is a lack of voice-enabled interfaces, forcing workers to rely on screens when hands-free operation is needed. Furthermore, reporting is often delayed until workers return to an office, leading to significant data loss and reduced accuracy in predictive maintenance insights.

Another major pain point is the ineffective integration of maintenance documentation with AI systems. Current solutions cannot dynamically access factory-specific knowledge, such as Original Equipment Manufacturer (OEM) handbooks, standard operating procedures, and maintenance logs. Text-based interfaces are inadequate for the immediate, contextual needs of shop floor conditions. Consequently, valuable technical knowledge remains inaccessible to non-technical operators during critical maintenance tasks, limiting the adoption and effectiveness of advanced manufacturing assistance solutions.

Objective 1: 

The primary objective is to enhance maintenance operations by transforming complex technical data into actionable guidance. This involves integrating voice-enabled AI with simulated real-time sensor data and documentation using the OpenVoiceOS framework. By doing so, the experiment aims to provide manufacturers with intelligent digital assistants that streamline maintenance workflows and improve decision-making on the shop floor.

Objective 2:

The experiment seeks to significantly improve accessibility for non-technical operators. By developing advanced AI capabilities that are accessible via intuitive voice and text interfaces, the project aims to lower the barrier to entry for using sophisticated predictive maintenance tools. This ensures that operators can interact naturally with the system without needing specialized data science expertise.

Objective 3: 

A further objective is to establish a continuous learning cycle through a RAG-LLM architecture. The system is designed to incorporate operator feedback loops, allowing the AI models to refine their accuracy and relevance over time based on human input. This objective creates a scalable solution that adapts to specific manufacturing environments and improves its troubleshooting capabilities through actual usage.

The experiment targets the manufacturing sector, specifically focusing on the home appliances industry and broader maintenance management applications. The use case context involves the “G45 cavity production line” at the BEKO Europe facility in Cassinetta, Italy. This setting serves as the industrial validation environment where the system will be tested against the rigors of discrete manufacturing processes.

The operational environment for this specific experiment utilizes a simulation-based approach using historical production and maintenance data provided by BEKO Europe. The target users are maintenance managers, engineers, and shop floor operators who require immediate assistance with equipment anomalies. Relevant constraints include the need for strict adherence to data privacy regulations (GDPR and AI Act) regarding voice data, and the technical requirement to integrate with existing legacy documentation and machine-extracted data logs.

The experiment addresses several operational and technical challenges related to the quality control of second-life batteries used in the Wall-AI. The manufacturing process involves the disassembly of electric vehicle batteries and the reassembly of selected modules into new CAVE battery packs. While this approach supports circular economy and sustainability goals, it introduces complexity in ensuring consistent quality, safety, and performance.

One of the main challenges is the lack of full digital traceability throughout the disassembly and assembly stages. Currently, information related to battery module testing, classification, and integration may be recorded manually. This can limit the ability to track each module’s history, monitor its condition over time, and ensure complete transparency across the production workflow.

Another important issue is the dependence on manual data entry and subjective evaluations during inspection and classification. Technicians must assess parameters such as voltage behavior, remaining capacity, and physical condition. While their expertise is critical, manual processes can introduce variability and inconsistencies between operators.

There is also limited real-time visibility of quality metrics and test results across the manufacturing line. Without centralized and structured digital monitoring, it becomes more difficult to detect anomalies quickly or compare performance data between batches. This reduces the ability to react proactively to emerging issues and may delay corrective actions.

Finally, identifying early trends or deviations that could affect long-term battery performance remains a challenge. Reused battery modules may show gradual performance changes due to aging or prior usage conditions. Without systematic data analysis and structured monitoring, subtle patterns may go unnoticed until they affect system reliability.

The experiment therefore focuses on improving digitalization, standardization, and data-driven quality control within the Wall-AI manufacturing process.

Objective 1: 

Enhancement of product quality:

Integrating the Digital Assistant (DA) into the testing process enables precise, consistent, and automated quality control of each Wall-AI battery. This integration ensures standardized testing procedures, minimizes variability, and improves traceability. As a result, product reliability and long-term performance are strengthened, enhancing customer satisfaction and reinforcing EAVE’s reputation as a provider of high-quality, sustainable energy solutions.

Objective 2:

Competitiveness and cost reduction:

The incorporation of digital assistance into the testing workflow increases operational efficiency by providing operators with clear, standardized guidance and automated data management. This reduces process time, minimizes errors and rework, and optimizes resource utilization. Consequently, production costs are lowered, and EAVE’s competitive position in the energy storage market is reinforced.

Objective 3: 

Labour security and process effectiveness:

Although operators remain responsible for all physical testing tasks, the Digital Assistant supports them by delivering step-by-step guidance, real-time instructions, and consistent enforcement of safety protocols. This support enhances workplace safety, reduces the likelihood of human error, and improves overall process reliability and effectiveness.

The experiment is positioned within the renewable energy sector, specifically in the field of solar energy integration and second-life battery reuse. It focuses on the application of OVOS to support the quality control of Wall-AI battery manufacturing. Wall-AI is an advanced energy management device designed to optimize energy storage and consumption in self-consumption photovoltaic systems, helping users reduce electricity costs and improve system efficiency.

The experiment will be carried out at the Institute of Technology and Renewable Energy (ITER) in Tenerife, Spain. It will take place in a real industrial production environment where Wall-AI batteries are disassembled, tested, reassembled, and validated. The Digital Assistance solution will therefore be tested under operational manufacturing conditions rather than in a laboratory-only setting.

The main users involved are battery manufacturing quality control operators. Key stakeholders include renewable energy system operators and future end users of Wall-AI systems.

The implementation of the Digital Assistant must comply with data protection regulations, as limited operational and user interaction data may be processed. The system is designed to comply with the General Data Protection Regulation (GDPR) by minimizing personal data collection and focusing primarily on process-related information.

Information Fragmentation and Accessibility

Operationally, the warehouse workforce faces a practical “information disconnect” that creates noticeable inefficiencies in daily logistics. Frontline workers handling physical goods usually lack direct access to digital systems while moving through the facility. To answer routine logistical or technical questions, such as “Is article X currently in stock?”, “Is this timber batch treated for outdoor use?”, or “Where is commission Y stored?” they frequently need to interrupt their workflow to walk to a central desktop terminal. Alternatively, they must locate and ask a more experienced colleague. This manual information retrieval causes recurring disruptions and unnecessarily ties up the resources of senior staff. Ultimately, the lack of immediate, hands-free data access at the point of action slows down the picking process and leads to avoidable delays in overall warehouse operations.

Operational Inefficiencies in Logistics

From a logistical perspective, the lack of real-time, hands-free information access creates bottlenecks. Warehouse staff often struggle to locate packed commissions or verify stock levels instantly while operating machinery or handling goods. The current environment lacks a digital interface that supports the mobile, hands-busy nature of timber logistics, leading to avoidable search times, picking errors, and delayed order processing.

Acoustic and Hardware Constraints for Speech-to-Text

Beyond operational hurdles, establishing a reliable voice interface in this specific setting presents a massive technical challenge. The timber yard is inherently loud, characterized by the continuous noise of heavy forklifts, machinery, and the physical handling of massive wooden goods. Furthermore, the workforce relies on standard two-way radios for mobile communication. These devices transmit highly compressed, narrowband audio that frequently suffers from static, frequency interference, and clipped sentences (push-to-talk delays). Achieving accurate Speech-to-Text (STT) recognition and extracting exact data points, such as alphanumeric article numbers or matchcodes, from such noise-polluted, low-quality audio streams is highly complex. Standard STT models struggle under these harsh conditions, requiring the underlying AI system to be exceptionally robust in processing garbled input to ensure the assistant remains a helpful tool rather than a source of frustration.

Objective 1: Enable Real-Time, Hands-Free ERP Interaction

The first and main objective is to bridge the gap between the physical workforce and digital record-keeping systems. The experiment aims to develop and deploy “Agent” capabilities within the Open Voice OS (OVOS) framework that allow the DIA to autonomously query the ERP and WMS via APIs. This will enable logistics staff to perform stock checks, track order statuses, and locate packed commissions using voice commands via radios, targeting a 40% reduction in average information retrieval times.

Objective 2: Validate User Acceptance in a Non-Desk Environment

The second objective is to successfully integrate the conversational AI into the daily routine of the logistics workforce. The experiment seeks to demonstrate that a voice-first interface via radios is a highly viable and preferred tool for a rough, hands-busy industrial setting. A key success metric is achieving an 80% user adoption rate among the warehouse staff within the first two months, proving that the solution effectively reduces their cognitive load without disrupting physical workflows.

Objective 3: Democratize Technical Knowledge via RAG

The third objective is to make complex, unstructured technical knowledge instantly accessible to all operational employees, regardless of their tenure. By implementing a Retrieval-Augmented Generation (RAG) pipeline, the experiment aims to ingest and index vast libraries of supplier documents and internal guidelines. The goal is to enable the Digital Intelligent Assistant (DIA) to answer specific technical queries with high accuracy (target: 90%), thereby reducing the dependency on senior experts and accelerating the autonomy of junior warehouse staff.

Sector and Operational Environment

The experiment takes place in the Timber Trade and Logistics sector. To validate the solution in a real-world scenario, the pilot is carried out in an operational test environment provided by Holzhandel Vogt in Germany, which acts strictly as the use-case provider. The setting focuses specifically on the active warehouse and logistics yard, where goods are actively picked, packed, and loaded. This rough industrial environment is characterized by high noise levels, heavy machinery such as forklifts, and a vast inventory of varied physical products, necessitating robust and hands-free communication methods.

Target Users and Stakeholders

The primary target users are warehouse employees and logistics staff who require immediate, hands-free information access to maintain safety and efficiency during their daily physical operations. By utilizing two-way radios, these workers can retrieve essential operational data without interrupting their physical tasks. Secondary stakeholders include the IT department of the use-case provider, which manages the underlying ERP and Warehouse Management Systems (WMS), as well as the management team, which is highly interested in workforce resilience, knowledge preservation, and process optimization.

Constraints and Requirements

The experiment must adhere to strict data privacy and security standards, including GDPR, particularly regarding any employee data or internal operational metrics processed by the LLM. From a technical perspective, the solution must seamlessly integrate with the existing ERP and other system infrastructure via REST APIs and MQTT, and function reliably through the audio interface of two-way radios operating within the warehouse’s existing communication network. Safety is paramount in this setting; the voice interface must provide clear, concise information without distracting operators while they are maneuvering heavy machinery or handling materials.

Being a part of the automotive industry, we are subject to a range of legal and regulatory frameworks relating to labor law, workplace safety, data protection, product quality, and constructural compliance within the automotive standards. As a European SME, we operate in accordance with national and European labour legislation, which covers everything from recruitment to induction, equal treatment, skills development and employee rights, and obliges us to provide adequate and verifiable training, including safety instructions, before employees take on specific production tasks.

These legal aspects impact the design, operation and deployment of the proposed digital solution. Therefore, the trial must be aligned with applicable legal obligations to ensure safe, ethical and compliant implementation.

The onboarding process relied primarily on traditional training, a hands-on knowledge transfer, leading to creating variability in training quality. Mostly informal, with senior workers transmitting skills through demonstrations and written/oral instructions. Training materials and communication are often in the native language, which can be a barrier for non-native speakers. This often resulted in errors, inconsistent delivery, material loss and increased workload.

One of our primary objectives was to enhance workforce readiness by simulating our real manufacturing environment to effectively prepare newly hired employees for actual production operations before they enter the production floor.

Objective 1: 

Design and deploy an interactive digital training platform utilizing Conversational AI and HMI that mimics real-world manufacturing tasks, focusing on plastic injection molding and wire harness assembly. The created virtual platform will allow trainees to practice procedures, troubleshoot scenarios, and build confidence in a safe and controlled environment.

Objective 2:

Create a training module in VR, integrating visual tools such as step-by-step video tutorials, incorporating gamified solutions that reward accuracy, speed, and adherence to standard operating procedures.

Objective 3: 

Implement a feedback mechanism through AI-driven analytics to track trainee performance and adapt training content in real-time. These tools will also help measure the effectiveness of the training platform and identify areas for improvement. Performance metrics such as task completion time, accuracy, and safety in the testing environment will be continuously monitored through simulation (VR cast) on a tablet device.

The experiment will be conducted at ELVEZ, a manufacturing SME based in Višnja Gora, Slovenia, which specializes in producing high‑quality automotive components. The company operates two main production lines: injection molding of complex plastic parts and assembly of cable harnesses. The initial phase of the experiment focuses on a work procedure within the injection‑molding production line, where operators are responsible for assembling intricate plastic components. In a later phase, the experiment will expand to include the development of a new training program dedicated to the production of cable harnesses and assemblies.

The experiment will take place directly on the injection molding production floor at ELVEZ, representing a real industrial environment rather than a laboratory or pilot setting. This environment includes multiple injection molding machines, welding and assembly equipment, and testing stations used for quality control. The goal is to evaluate the solution under real production conditions (TRL6), where operators must work with precision and adhere to established procedures. On the production floor, we have established a dedicated clean and controlled area specifically prepared for safe VR‑based training. This space ensures that operators can participate in immersive training sessions without interfering with ongoing production activities and while maintaining all required safety standards. Before participating in the training, individuals must complete a questionnaire specifically designed to determine whether they are suitable for using VR technology. This assessment helps ensure that participants can safely and comfortably engage with immersive training tools. In addition, each individual participant is given the opportunity to choose whether they prefer to take part in the VR-based training program or follow the traditional training approach, allowing them to choose what method best aligns with their comfort level and learning preferences.

The use case is based on a multi‑step procedure in which an operator is required to produce a fully functional injection‑molded plastic part. The component is manufactured in two separate pieces that must be joined through a precise welding process. Operators must follow detailed instructions and perform several critical tasks, such as correctly positioning parts in the cradle, executing the welding operation, testing the components for tightness, inspecting them for defects, and marking and documenting the finished parts. Common errors observed among operators include incorrect placement of parts, insufficient or omitted tightness testing, failure to identify defects in components or assemblies, and incorrect marking or documentation, which can lead to quality‑control issues.

The primary users involved in the experiment are production operators, particularly newly hired employees who often face challenges due to the complexity of the processes. Additional stakeholders include shift leaders and supervisors who oversee daily operations, quality assurance personnel responsible for maintaining product standards, process engineers who design and optimize workflows, and company management. Customers in the automotive supply chain also indirectly benefit from improved process reliability and product quality.

The experiment must comply with several operational and regulatory constraints. These include adherence to automotive industry safety and quality standards such as IATF 16949, compliance with machine safety regulations for working near injection molding and welding equipment, and meeting certification requirements for operators performing specialized tasks.

Access to production data may be restricted due to proprietary or customer‑sensitive information, and the solution must integrate with existing digital systems such as manufacturing execution systems and quality‑tracking tools. Additionally, production lines have limited availability for downtime, which requires careful planning to avoid disruptions.