Depth information is crucial for many modern computer vision applications, yet common low-end depth sensors struggle to meet algorithmic demands due to poor spatial resolution, noise, and information loss. This is especially problematic in high-precision fields like industrial robotics, where millimeter-scale accuracy is critical. To address these real-world limitations, we present a comprehensive depth enhancement approach that combines depth super-resolution for improved spatial detail and depth completion for recovering missing data. We build upon the SGNet model for super-resolution, significantly enhancing it by cleverly applying loss functions during training and integrating functionalities from DepthAnythingV2, a state-of-the-art transformer-based architecture for monocular depth estimation. Our method achieves a substantial improvement over the considered baseline, as validated through extensive experiments on the NYUv2 benchmark, the MetaGraspNet dataset, and a challenging real-world internal dataset. This work highlights a robust solution for acquiring high-quality depth data, pushing the boundaries of what low-cost sensors can achieve.

Object detection plays a pivotal role in industrial automation, where systems must frequently adapt to novel scenarios with minimal configuration overhead. Conventional approaches require extensive data collection, annotation, and retraining, making them ill-suited for dynamic, real-time environments. In this work, we present a plug-and-play detection pipeline based on a state-of-the-art image-conditioned transformer network, OWLv2, tailored for industrial applications. Our framework enables the integration of new object categories without retraining or server interruption, offering flexible and user-controllable parameters for fine-tuning detection behavior. We demonstrate the pipeline’s versatility across four key industrial use cases: bin picking, anomaly detection, depalletizing, and logo verification. The system is designed for real-time deployment via a robust server-client architecture and achieves competitive performance on standard benchmarks while maintaining adaptability and efficiency in operational settings. The source code of this work is available at GitHub (https://github.com/milenayahya/OneShotObjectDetection), where also some videos are available.

In this paper, we present an advanced control framework for power management applications, named Neural Adaptive Model Predictive Control (NA-MPC), designed to provide an optimal power allocation among multiple energy sources, perform a multi-objective online adaptation of the optimal control policy, and ensure a fast real-time execution with low computational demand. NA-MPC augments general MPC problems with three key features: 1) an online metaheuristic tuning strategy adapts the MPC cost function weights, to attain multiple concurrent control objectives at once; 2) through neural emulation, the MPC control policy is replaced by an equivalent neural MPC controller, exhibiting universal approximation guarantees and ensuring real-time feasibility; 3) a neural black-box MPC prediction model is employed, identified only via noise-corrupted input-output measurements from the plant, which is assumed to be unknown. The general formulation and versatility of NA-MPC make it potentially applicable to several power management scenarios; in this work, we apply NA-MPC to the case study of power management in fuel cell hybrid electric vehicles (FCHEVs), a topic of growing interest within the frame of sustainable transportation, for which novel and efficient strategies are still lacking. The effectiveness of NA-MPC is thoroughly assessed via numerical simulations, demonstrating its capability to optimally attain multiple control objectives concurrently in real time; moreover, NA-MPC consistently outperforms the most prominent state-of-the-art HEV power management strategies. Note to Practitioners—The aim of this paper is to introduce an advanced online-adaptive optimal control strategy, named NA-MPC, and employ it as a novel power management strategy for FCHEVs, with the purpose of addressing several technical shortcomings of the existing state-of-the-art strategies. Specifically, the latter typically fail in performing effective trade-offs between accurate power tracking and supply consumption, proving a merely suboptimal control action. Such strategies have also very limited adaptation capabilities, being either offline-tuned or employing simple non-optimal adaptation policies. Moreover, only few basic optimal control strategies are proposed in the literature, with little focus on their real-time feasibility. By contrast, our NA-MPC strategy provides an optimal power allocation, effectively attains multiple concurrent control objectives, and, thanks to its neural embedding, is real-time feasible and easily implementable on hardware with limited computational resources. Furthermore, the general formulation and versatility of NA-MPC enable its potential application across a wide variety of different power management scenarios.

This paper addresses the distributed localization problem for a network of sensors placed in a three-dimensional space, in which sensors are able to perform range measurements, i.e., measure the relative distance between them, and exchange information on a network structure. While most existing studies primarily develop localization algorithms under the assumption that the entire sensor network is localizable, the problem of determining whether a sensor is localizable has received limited attention. However, neglecting this preliminary step can significantly hamper the accuracy of localization algorithms due to error propagation from unlocalizable sensors in iterative localization procedures. To address this research gap, we focus on two key challenges: i) deriving rigorous theoretical results and developing algorithms to verify sensor localizability, and ii)designing an efficient distributed localization algorithm that utilizes these localizability results. Specifically, we start by deriving a necessary and sufficient condition for sensor localizability using barycentric coordinates. Then, building on this theoretical result, we design a distributed localizability verification algorithm, in which we propose and employ a novel distributed finite-time algorithm for sum consensus. Finally, we develop a distributed localization algorithm based on conjugate gradient method and derive theoretical guarantees on its performance, ensuring finite-time convergence. The efficiency of our algorithm compared to the existing ones from the literature and its capability to handle scenarios with moderate levels of noise in the measurements are further demonstrated through numerical simulations.

Smooth and collision-free trajectory planning is crucial to high speed and high precision machining, such as 3D laser cutting. However, it is difficult to further enhance the kinematic performance of the primary translational axes during the process. This paper presents a novel two-phase planning strategy, which optimizes the tool orientation and leverages a redundant standoff axis to significantly enhance the smoothness of the translational movements in redundant 3D laser cutting machines. In the first phase, collision-free configuration spaces (C-spaces) are constructed along the tool path, utilizing a graph-based search approach with Dijkstra's algorithm for tool orientation optimization. Subsequently, a secondary orientation curve, namely the M path, is planned in the second phase with a variable distance from the primary tool path curve, and the motion of the redundant standoff axis is handled via a deep reinforcement learning approach. The proposed methodology provides an advancement in conventional five-axis machines lacking of flexibility. Experimental validation confirms the potential of the approach to substantially improve machining accuracy and efficiency.

In this paper, we propose a transformer-based architecture for predicting contact forces during a physical human-robot interaction. Our Neural Network is composed of two main parts: a Multi-Layer Perceptron called Transducer and a Transformer. The former estimates, based on the kinematic data from a motion capture suit, the current contact forces. The latter predicts – taking as input the same kinematic data and the output of the Transducer – the human motions and the contact forces over a time window in the future. We validated our approach by testing the network on directions of motions that were not provided in the training set. We also compared our approach to a purely Transformer-based network, showing a better prediction accuracy of the contact forces.

In this paper we consider the cooperative control of the manipulation of a load on a plane by a team of mobile robots. We propose two different novel solutions. The first is a controller which ensures exact tracking of the load twist. This controller is partially decentralized since, locally, it does not rely on the state of all the robots but needs only to know the system parameters and load twist. Then we propose a fully decentralized controller that differs from the first one for the use of i) a decentralized estimation of the parameters and twist of the load based only on local measurements of the velocity of the contact points and ii) a discontinuous robustification term in the control law. The second controller ensures a practical stabilization of the twist in presence of estimation errors. The theoretical results are finally corroborated with a simulation campaign evaluating different manipulation settings.

In this paper, a distributed approach for the estimation of kinematic and inertial parameters of an unknown rigid body is presented. The body is manipulated by a pool of ground mobile manipulators. Each robot retrieves a noisy measurement of its velocity and the contact forces applied to the body. Kinematics and dynamics arguments are used to distributively estimate the relative positions of the contact points. Subsequently, distributed estimation filters and nonlinear observers are used to estimate the body mass, the relative position between its geometric center and its center of mass, and its moment of inertia. The manipulation strategy is functional to the estimation process, and is suitably designed to satisfy nonlinear observability conditions that are necessary for the success of the estimation. Numerical results corroborate our theoretical findings.

In this work a nonlinear dynamical model of a fuel cell stack is developed by means of artificial neural networks. The model presented is a black-box model, based on a set of easily measurable exogenous inputs like pressures and temperatures at the stack and is able to predict the output voltage of the fuel cell stack. The model obtained is being exploited as a component of complex control systems able to manage the energy flows between fuel cell stack, battery pack, auxiliary systems and electric engine in a zero-emission vehicle prototype

This work deals with the discrimination of neutrons and gamma rays on the basis of their different pulse shapes in scintillator detectors; this technique is widely employed in nuclear fusion applications. After a thorough phase of data analysis, a multi layer perceptron (MLP) is trained with the aim of processing the shape of light pulses produced by these ionizing particles in an organic liquid scintillator and digitally acquired. Moreover, fast-superimposed events (called pile-ups) are detected and a further MLP is trained to analyze them and recover the original superimposed events. Satisfactory experimental results were obtained at the Frascati Tokamak Upgrade, ENEA-Frascati, Italy.

In this paper it is shown that, with slight modifications, State Controlled CNNs (SC-CNNs) are able to approximate the behaviour of a class of complex dynamics with multivariable nonlinearities. In particular, in the so-called Extended SC-CNN defined in this work, the output nonlinearity shape has been modified, and a new template acting on the output function of the cell has been introduced. The needed circuitry to extend SC-CNNs, together with SPICE simulations of the new system, are here reported in order to confirm the suitability of the approach.