Gašper Škulj

We present Century Maze,1 a simulated environment for cooperative navigation of groups numbering over 100 robots with primarily visual observations. The underlying task of large-scale object search is designed to push the scalability and efficiency of decentralized multi-robot systems relying on communication to overcome the limitations of separate units. To enable the training of robotic agents in large-scale visual embodied applications such as Century Maze, we propose two reinforcement learning (RL) components that facilitate the emergence of effective communication: Differentiable Inter-Agent Belief Learning (DIABL), which provides a clear learning signal by optimizing the agents' beliefs about their goals and propagating gradients back over a differentiable communication channel to include the appropriate observers, and Informative Event Replay (InfER), which compensates for the scarcity of samples relevant to DIABL by maintaining a dedicated buffer for auxiliary learning. We demonstrate our combined method through an actor-critic neural network architecture encompassing a decentralized multi-agent system of variable size with individual goal-based belief modules. Results on Century Maze show that our method is able to significantly improve performance through cooperation between over 100 agents against comparable baselines and maintains its effectiveness in realistic conditions.

Deploying deep learning models on low-power devices is an increasingly important research challenge. Microcontrollers, in particular, have strict limitations due to the limited processing power and memory. Specifically, object localisation on such devices remains in the early stages of research, showing the need for architectural designs tailored to resource-constrained environments. In this work, we implement and evaluate different architectural configurations of deep learning models on a miniature mobile robot operating in a granular, sand-like terrain, where both robots and rocks must be localised as a prerequisite for robust navigation. The localisation models are based on the MobileNetV2 architecture utilising the FOMO (Faster Objects, More Objects) algorithm with dynamic convolutions. The proposed system consists of an HM01B0 greyscale camera integrated with a low-power RP2040 microcontroller. The object detection models are trained on a custom dataset consisting of 10,000 greyscale images (160×160 px) featuring robots and rocks. Our findings indicate that the best trade-offs between accuracy and memory are achieved by truncating the middle layers of the base model, balancing output resolution with retained pre-trained features and classifier complexity. Additionally, the results indicate that increasing the input image resolution produces a greater improvement in accuracy compared to expanding the model’s classifier head, relative to the increase in inference time. This study presents the first large-scale variation of FOMO truncations for the purpose of object localisation on low-power microcontrollers and offers practical guidance for selecting neural network architectures.

Operation and maintenance (O&M) events resulting from environmental factors (e.g., precipitation, temperature, seasonality, and unexpected weather conditions) are among the primary sources of operating costs and downtime in run-of-river small hydropower plants (SHPs). This paper presents a data-driven methodology for predicting such long events using machine learning models trained on historical power production, weather radar, and forecast data. Case studies on two Slovenian SHPs with different structural designs and levels of automation demonstrate how environmental features—such as day of year, rain duration, cumulative amount of rain, and rolling precipitation sums—can be used to forecast long events or shutdowns. The proposed approach integrates probabilistic classification outputs with threshold-consistency smoothing to reduce noise and stabilize predictions. Several algorithms were tested—including Logistic Regression, Support Vector Machine (SVM), Random Forest, Gradient Boosting, and k-Nearest Neighbors (k-NN)—across varying feature combinations for O&M model development, with cross-validation ensuring robust evaluation. The models achieved an F1-score of up to 0.58 in SHP1 (k-NN), showing strong seasonality dependence, and up to 0.68 in SHP2 (Gradient Boosting). For SHP1, the best model (k-NN) correctly detected 36 long events, while 15 were misclassified as no events and 38 false alarms were produced. For SHP2, the best model (Gradient Boosting) correctly detected 69 long events, misclassified 23 as no events, and produced 42 false alarms. The findings highlight that probabilistic machine learning-based forecasting can effectively support predictive O&M planning, particularly for manually operated or service-operated SHPs.

This paper presents a novel approach to thermal process control in the food industry, specifically targeting the pasteurization and cooking of soft-boiled eggs. The unique challenge of this process lies in the precise temperature control required, as pasteurization and cooking must occur within a narrow temperature range. Traditional control methods, such as fuzzy logic controllers, have proven insufficient due to their limitations in handling varying loads and environmental conditions. To address these challenges, we propose the integration of robust reinforcement learning (RL) techniques, particularly the utilization of the Deep Q-Network (DQN) algorithm. Our approach involves training an RL agent in a simulated environment to manage the thermal process with high accuracy. The RL-based system adapts to different heat capacities, initial conditions, and environmental variations, demonstrating superior performance over traditional methods. Experimental results indicate that the RL-based controller significantly improves temperature regulation accuracy, ensuring consistent pasteurization and cooking quality. This study opens new avenues for the application of artificial intelligence in industrial food processing, highlighting the potential for RL algorithms to enhance process control and efficiency.

This paper addresses the problem of establishing efficient intralogistic systems, focusing on the generation of roadmaps on a given layout and the coordination of multiple Automated Guided Vehicles (AGVs). A simulation-based approach for automatic roadmap design is proposed. An event-based simulator is developed that uses ant-colony inspired optimization to generate roadmaps tailored to the specific characteristics of a given intralogistic problem, i.e., the plant layout, fleet size, statistical description of tasks, dispatching algorithm, etc. The generated solutions are evaluated with a Multi-Agent Path Finding (MAPF) simulator that uses a Safe Interval Path Planning (SIPP) algorithm. By analysing the system throughput, the optimal fleet size for the system is proposed. The approach is validated through various examples and benchmarked against existing methods in the literature. Note to Practitioners —Creating efficient roadmaps plays a key role in improving the efficiency of multi-AGV systems. Usually, roadmaps are designed by experts, which is time consuming and may lead to suboptimal solutions due to the complexity of the problem. The main objective of this work is to develop a methodology that allows the automatic generation of efficient roadmaps taking into account the specific characteristics of an intralogistic problem. The main advantage is the fast generation of solutions and their validation in terms of throughput. At the same time, the presented methodology allows a quick correction of the solution if one of the problem characteristics changes, without the need to change the layout itself. This provides industry practitioners with an effective tool that facilitates their work in setting up multi-AGV systems, thus saving time and costs.

Miniature mobile robots in multi-robotic systems require reliable environmental perception for successful navigation, especially when operating in the real-world environment. One of the sensors that have recently become accessible in microrobotics due to their size and cost-effectiveness is a multi-zone time-of-flight (ToF) sensor. In this research, object classification using a convolutional neural network (CNN) based on an ultra-low resolution ToF sensor is implemented on a miniature mobile robot to distinguish the robot from other objects. The main contribution of this work is an accurate classification system implemented on low resolution, low processing power and low power consumption hardware. The developed system consists of a VL53L5CX ToF sensor with an 8x8 depth image and a low-power RP2040 microcontroller. The classification system is based on a customised CNN architecture to determine the presence of a miniature mobile robot within the observed terrain, primarily characterized by sand and rocks. The developed system trained on a custom dataset can detect a mobile robot with an accuracy of 91.8% when deployed on a microcontroller. The model implementation requires 7 kB of RAM, has an inference time of 34 ms, and an energy consumption during inference of 3.685 mJ.

This study investigates the effects of blockchain technology scalability limitations on the performance of Blockchain-Based Shared Manufacturing (BBSM), an innovative smart-manufacturing paradigm aimed at enhancing the utilization of global manufacturing resources via peer-to-peer (P2P) collaboration of self-organized manufacturing assets. Despite the prevalence of research highlighting blockchain technology’s scalability limitations as the main barrier for adoption, few studies have explored their effects on the operation of blockchain-based systems. The primary goal of the presented research work is to explore the implications of blockchain technology scalability limitations on the BBSM system’s performance and user behavior. To obtain realistic behavior, an experiment is conducted using an online game played by human participants. Analysis of the players’ strategy is used for implementation of a multi-agent simulation model, which is then employed to assess the influence of varying blockchain network configurations on the BBSM concept’s performance. Preliminary experimental findings reveal that a congested blockchain network leads to increased transaction costs and reduced service prices, consequently devaluing the manufacturing role in the BBSM system and causing underutilization of existing maximum production capacities. Moreover, allocating funds to financial activities rather than manufacturing activities yields superior outcomes for system users. Simulation results indicate that the BBSM system’s response to alterations in blockchain network throughput is contingent upon the production function. The findings of this study reveal that the scalability limitations of blockchain technology impair the performance of the BBSM system and affect user behavior in the system, underscoring the necessity for future research to concentrate on incorporating scalable solutions within blockchain-based manufacturing systems.

Increasing the accessibility of collaborative robotics requires interfaces that support intuitive teleoperation. One possibility for an intuitive interface is offered by wearable systems that measure the operator’s movement and use the information for robot control. Such wearable systems should preserve the operator’s movement capabilities and, thus, their ability to flexibly operate in the workspace. This paper presents a novel wireless wearable system that uses only inertial measurement units (IMUs) to determine the orientation of the operator’s upper body parts. An algorithm was developed to transform the measured orientations to movement commands for an industrial collaborative robot. The algorithm includes a calibration procedure, which aligns the coordinate systems of all IMUs, the operator, and the robot, and the transformation of the operator’s relative hand motions to the movement of the robot’s end effector, which takes into account the operator’s orientation relative to the robot. The developed system is demonstrated with an example of an industrial application in which a workpiece needs to be inserted into a fixture. The robot’s motion is compared between the developed system and a standard robot controller. The results confirm that the developed system is intuitive, allows for flexible control, and is robust enough for use in industrial collaborative robotic applications.