Soft sensors are mathematical models of industrial processes that are often used for monitoring and control. Data-driven techniques based on artificial intelligence are generally used for identifying these models. Data scarcity is a challenging problem that occurs when the variable to be estimated must be measured by laboratory analysis. Here, a difference-based neural network called Δ -Net is proposed to develop dynamic nonlinear soft sensors when only a few hundred labeled data are available. The Δ -Net, which exploits the case difference heuristic approach, is based on two parallel neural networks responsible for processing pairs of input samples. The Δ -Net is trained on an augmented dataset obtained by pairwise ordering the small original dataset. The outputs of the subnets are then combined to estimate the difference between the corresponding outputs of the pairs. The reconstruction of the output sample is computed as the mean of the distances from the sample to a set of anchor points, to increase the robustness of the prediction. The proposed approach has been applied to well-known industrial benchmarking datasets. The obtained results show superior performance compared to other data augmentation approaches, including bootstrap resampling, variational autoencoders and Wasserstein generative adversarial networks.

In this work, black-box modeling techniques are applied to Bacterial Cellulose-based sensors to characterize their dynamic behavior. Several classes of linear and nonlinear models, including Finite Impulse Response, AutoRegressive with eXogenous Input, Nonlinear Finite Impulse Response, Nonlinear AutoRegressive with eXogenous Input, and Long Short-Term Memory networks, are developed and compared. The performance of each model is evaluated based on a one-step-ahead prediction and a ∞-step-ahead simulation using standard performance metrics such as Root Mean Square Error, Mean Absolute Error and the coefficient of determination. The results show the strengths and limitations of the different modeling approaches in capturing the dynamics of BC-based transducers. The ARX model showed the best results for the prediction in one step, but poor results were obtained when the prediction was considered in ∞ steps. The NFIR model is instead the best choice for long-term prediction.

Bacterial Cellulose (BC) functionalized with Ionic Liquids (ILs) is a promising candidate for sustainable electroactive sensors. While single-carrier transport models are well-established in piezoionic electroactive polymers, their applicability to BC-IL systems remains unverified. This study introduces a simplified 1D finite element model, significantly improving computational efficiency while preserving key physical insights. A detailed parametric analysis investigates the impact of different charge transport assumptions, revealing that single-carrier models are insufficient to fully describe the mechanoelectric transduction behavior. The results emphasize the necessity of a dual-carrier framework to accurately model BC-based transducers, offering a deeper understanding of multi-ionic interactions within the porous BC structure. By highlighting key mechanisms and limitations, this work provides a foundation for optimizing BC-IL sensors, preparing the way for more reliable and scalable bioelectronic applications.

In this work, a spiking neural network is developed that acts as a path encoder and processes a sequence of visual scenes of a realistic, highly unstructured environment. It is able to memorize a simulated four-legged robot and then guide it to a given goal. The overall structure, which exploits the functional roles of specific parts of the fruit fly brain, makes it possible to reproduce the same results as other existing structures while largely reducing the overall complexity of the network. This enables the real-time implementation of real robot navigation. The simulation results confirm the performance of the proposed spiking neural network in controlling a four-legged robot Unitree Go2 simulated in Gazebo.

In the field of robotics, ensuring reliable and efficient performance is crucial, especially when robots are entrusted with critical tasks. Anomaly detection systems play a crucial role in maintaining this reliability by detecting deviations from normal behavior and taking timely interventions. Traditional model-and knowledge-based approaches, while effective in controlled environments, reach their limits in dynamic and resource-constrained settings due to their reliance on predefined models, expert knowledge and high computational requirements. While data-driven methods, especially those using deep learning, offer better adaptability, they also bring challenges in terms of energy consumption and hardware limitations. To address these issues, this paper proposes a hybrid anomaly detection system that utilizes Spiking Neural Networks (SNNs) and Convolutional Neural Networks (CNNs). The SNN provides low-power processing of input features for anomaly detection, while the CNN efficiently extracts spatial features from high-resolution images to realize an accurate, high-performance classifier that can be used to perform an online fine-tuning of the SNN. A large dataset for robot navigation is used as a testbed. The obtained results show that this hybrid approach increases the speed and accuracy of anomaly detection while significantly reducing energy consumption, making it well-suited for applications in resource-constrained robotic systems.

Synchronization is one of the main studied behaviors in networks of oscillators. Particular synchronization manifolds are desired when designing Central Pattern Generators as a locomotion control strategy where, depending on the desired gait to implement, particular phase shifts must be imposed among the oscillators. Phase reduction theory provides a simple yet effective strategy to reduce the complexity of the problem to its essential elements. Thanks to the approximations of the piecewise linear (PWL) nonlinearity and slow-fast behavior of the FitzHugh-Nagumo (FHN) neural oscillator, harnessing singular perturbation analysis is possible to derive a closed-form solution of both the asymptotic phase and phase sensitivity function of the system, which is rarely possible. Using this approach, it is possible to relate the complex network of FHN neurons to a simple Kuramoto model, where synchronization conditions are already established.

This paper presents a novel circuit designed to directly connect neuromorphic dynamics with a motor system. The designed system is intended as an elementary building block for the construction of complex chains of adaptive neural oscillators that are directly controlled and in turn control the corresponding actuators. For the first time, the concept of a motor neuron as a unique neurocomputational and control unit is presented. This structure can be generalized to a broad class of actuators in which neuronal dynamics generate motion and the motor part, in turn, controls the neural dynamics. A new analog circuit that realizes a symbiotic relation between neural and motor activity is presented and the simulation results are reported and discussed.

Sleep disorders are continuously growing in the population and can have a significant negative impact on everyday life. Economic and non-invasive systems able to support the diagnosis procedure will be more and more adopted in the next years. The aim of this work is to investigate the classification performance of a convolutional neural network, based on a VGG structure, to identify obstructive sleep apnea events. A recently developed dataset containing audio signals recorded from high-quality contact microphones placed on the trachea of the subjects under study has been adopted to perform transfer learning over a pre-trained VGGish network. Spectrogram images have been extracted from the audio signals to serve as inputs for the classification process. The importance of the time window selection has been also investigated and comparisons with other recent methods proposed in the literature are reported.

The paper introduces a new structure of Liquid State Machine (LSM) characterised by local connectivity within the excitatory neurons of the reservoir layer. The architecture learning and testing capabilities are compared with the classical LSM network on an important robotic task of estimating exteroceptive information from proprioceptive signals coming from a simulated quadruped robot. The main advantages of the proposed architecture are discussed. The actual arrangement of the network resembles specific learning structures within the insect brain, endowed with interesting reaction-diffusion dynamics and classification capabilities. Simulation results are reported and compared with those ones obtained using a classic LSM. The proposed architecture, resembling typical biological solutions present in simple brains, through reaction-diffusion local mechanisms, significantly reduces the computational requirements needed by standard massively-connected neural network solutions. Moreover, the robustness of the introduced network, against faults in the sensory system, is demonstrated. In these conditions, the network is able to partially reconstruct the lacking sensory input signal from the learned relation with the other sensory input streams.

This paper aims to investigate the Liquid State Machines (LSMs) learning capability and robustness of a complex robot-environment interaction. The goal is to design an efficient robot state estimation method based on reservoir computing. The method maps local proprioceptive information acquired at the level of the leg joints of a simulated quadruped robot. The robot taken into account is the simulated version of Lilibot, a small-sized and reconfigurable bio-inspired robot with multiple real-time sensory feedback. Global information was provided from the ground reaction forces acquired on the tips of each leg. Simulation results are reported and compared, also in presence of faulty conditions in the sensory system.

This paper proposes the application of the Model Predictive Control (MPC) strategy to the locomotion of a quadrupedal robot endowed with a Central Pattern Generator (CPG) neural locomotion architecture. The neural structure is adaptive based on the proprioceptive information and the exteroceptive signals acquired through ground contact sensors. The MPC generates the high-level descending commands used by the CPG, controlling the robot navigation. Using its capability to provide optimized output guaranteeing, at the same time, the state and input constraints of the system, the MPC allows a robust heading control of the robot suitably interacting with the neural locomotion paradigm. The obtained results are analyzed and compared with those obtained in the same robotic architecture using a standard PID controller.

This paper aims to propose an innovative method to obtain the traversability maps of unstructured environments and the best path between two points on the basis of the specific characteristics of the robots that has to perform a given mission. Taken in consideration a robot team that have to traverse an assigned terrain, the peculiar capabilities of each robot are underlined in a dynamic simulation environment and then embedded into a neural network finally used as a robot model for the generation of the traversability maps. On the basis of the obtained results, the best robot within the team (wheeled, legged, hybrid) can be selected. The proposed strategy, together with the obtained simulation results, are presented, carefully analyzed and then compared.

In this paper a some examples of simulations and experiments performed in the last few years in the field of bio-inspired robotics are reviewed and revisited, deepening their characteristics and emphasising the role of imperfections that could be the main actors guiding their success in real environment. Our cases of study rely on both genetic and behavioral experiments on the fruit fly, from which models, simulations and robotic experiments were performed.

This manuscript aims at showing the effects of feature selection and manifold reduction methods in dealing with the wall-following problem in mobile robotics, a well-known nonlinearly separable classification problem in which sensor recordings are associated to controlled motor responses. The capabilities of state manifold reduction in Echo State Networks (ESNs) through Laplacian Eigenmaps (LEs) are described in terms of noise rejection over the trained weights. Furthermore, various machine learning-based and data mining-based methodologies are applied to show the advantages of using the most informative contents drawn from the original sensor readings.

Classification and sequence learning are relevant capabilities used by living beings to extract complex information from the environment for behavioral control. The insect world is full of examples where the presentation time of specific stimuli shapes the behavioral response. On the basis of previously developed neural models, inspired by Drosophila melanogaster, a new architecture for classification and sequence learning is here presented under the perspective of the Neural Reuse theory. Classification of relevant input stimuli is performed through resonant neurons, activated by the complex dynamics generated in a lattice of recurrent spiking neurons modeling the insect Mushroom Bodies neuropile. The network devoted to context formation is able to reconstruct the learned sequence and also to trace the subsequences present in the provided input. A sensitivity analysis to parameter variation and noise is reported. Experiments on a roving robot are reported to show the capabilities of the architecture used as a neural controller.

Sequence learning is a complex capability shown by living beings, able to extract information from the environment. Looking into the insect world, there are several examples where the presentation time of specific stimuli is considered to select the proper behavioural response. On the basis of previously developed neural models for sequence learning, inspired by the Drosophila melanogaster, a new formalization of key brain structures involved in the process is here provided. The input classification is performed through resonant neurons, stimulated by the complex dynamics generated in a lattice of recurrent spiking neurons modelling the Mushroom Bodies neuropile in the insect brain. The network devoted to the context formation is able to reconstruct the learned sequence and also to trace the subsequences present in the provided input. Simulation results were reported to show the capabilities of the architecture.

In this paper the implementation of a stable locomotion controller with sensory feedback on a hexapod robot structure is reported. Inspiration comes from recent results on the insect Drosophila melanogaster neural networks in charge for the control and modulation of basic crawling motion, where the role of sensory feedback is emphasized. A simple neural network, acting as a locomotion controller was designed and implemented. The phase stability, essential for a reliable gait generation, is assured exploiting tools from Partial contraction theory, whereas sensory feedback is used to locally modify the motor neuron dynamics to improve the robot dexterity in front of uneven terrains. Experimental results are reported in an autonomous hexapod robot, where the locomotion controller and sensory feedback are implemented in a commercial microcontroller low cost platform.

The aim of this paper is to propose a computational model, inspired by Drosophila melanogaster, able to handle problems related to motor learning. The role of the Mushroom Bodies and the Central Complex in solving this problem is analyzed and plausible biologically inspired models are proposed. The designed computational models have been evaluated in simulation using a dynamic structure inspired by the fruit fly. The obtained results open the way to new neurobiological experiments focused to better understand the underlined mechanisms involved, to verify the feasibility of the hypotheses formulated and the significance of the obtained results.

The concept of peripersonal space is an interesting research topics for psychologists, neurobiologists and for robotic applications. A living being can learn the representation of its own body to take the correct behavioral decision when interacting with the world. To transfer these important learning mechanisms on bio-robots, simple and efficient solutions can be found in the insect world. In this paper a neural-based model for body-size learning is proposed taking into account the results obtained in experiments with fruit flies. Simulations and experimental results on a roving platform are reported and compared with the biological counterpart.

Neural centers devoted to spatial memory and path integration were largely studied in rats and in different insect species like ants and bees. In this paper a neural-based model for the formation of a spatial working memory is proposed mirroring some peculiarities of the Drosophila central brain and in particular the ellipsoid body. Simulation results are reported opening the way to applications on roving platforms.

This paper proposes a neural-based model for ego-motion detection and compensation in applications related to bionic antennae. Touch or near range sensors are used on moving platforms to detect the presence of surrounding objects: additional feature could also be extracted, like distance, material characteristics and the shape of objects. The processing and control architecture is based on spiking neurons and in particular a series of resonate and fire neurons has been used to extract important features from the sensory data. The use of spiking resonant neuron arrays is treated as a general methodology for bio-inspired feature clustering. The use of the resonance allows to emphasize particular signal contents which depend on ego-motion: the detection and compensation of such components hidden in the acquired signal is a real added value towards the construction of robust and bio-inspired methodologies for simple and efficient forward models. The methodology is applied to a bionic antenna which was mounted on a roving robot. This could be very useful in different scenarios enriching the multimodal sensory system usually available on a navigation platform. Simulation results and experiments performed on a roving platform are reported.

Insects show advanced capabilities and a rich behavioral repertoire in task solving and thus they are becoming a reference point in Neuroscience for studying simple cognitive structures. In particular, thanks to many neurogenetic tools, the fruit fly Drosophila melanogaster became a relevant source of inspiration for Robotics. Mushroom Bodies (MBs) are very interesting neural structures involved in the regulation of behaviors in fruit fly, even if their main role regards olfactory conditioning. In this paper a novel bio-inspired neural architecture is presented, where the MBs role in attention tasks is focused. The model is a multi-layer spiking neural network where the MBs and their direct and or indirect interactions to other key elements of the insect brain, the Central Complex and the Lateral Horns, are modeled. The biological background of the proposed model is presented together with a detailed description of the architecture; simulation results and remarks on the biological counterpart are also reported.

The aim of this paper is to study the emergence of coordinated activities, and the investigation of collaboration between individuals in a small group of robots. The idea is to impose very simple global rules and to give a primary role to the environment mediation. In the paper the specialization strategy, already introduced in a previous work is extended, to autonomously solve a task assignment problem among agents in an initially homogeneous swarm. In particular, a given sequence of tasks is assigned to the group and each robot has to autonomously specialise in solving sub-sequences, resulting in a labor division which improves the performance of the team. Behavioral improvement is guided by a global reward function. Results, obtained in a dynamic simulation environment, show that performances depend by environmental conditions and starting positions of the singular agents: environment and the other robots play clearly a fundamental role in mediating the swarm capabilities.

In spite of their small brain, insects show a complex behavior repertoire and are becoming a reference point in neuroscience and robotics. In particular, it is very interesting to analyze how biological reaction-diffusion systems are able to codify sensorial information with the addition of learning capabilities. In this paper we propose a new model of the olfactory system of the fruit fly Drosophila melanogaster. The architecture is a multi-layer spiking neural network, inspired by the structures of the insect brain mainly involved in the olfactory conditioning, namely the Mushroom Bodies, the Lateral Horns and the Antennal Lobes. The Antennal Lobes model is based on a competitive topology that transduces the sensorial information into a pattern, projecting such information to the Mushroom Bodies model. This model is based on a first and second order reaction-diffusion paradigm that leads to a spontaneous emerging of clusters. The Lateral Horns have been modeled as an input-triggered resetting system. The structure, besides showing the already known capabilities of associative learning, via a bottom-up processing, is also able to realize a top-down modulation at the input level, in order to implement an expectation-based filtering of the sensorial inputs.

In this paper we investigate a novel algorithm for solving classification problems in an action-oriented perception framework supported by visual feedback. The approach is based on an extension of the Neural Gas with local Principal Component Analysis (NGPCA) algorithm. As an abstract Recurrent Neural Network (RNN) this model is able to complete a partially given pattern. Under this point of view it is possible to generalize the model as a supervised classifier in which for a given segmented object (i.e. with particular visual cues) the class variable is retrieved as the network outputs. An incremental version of the algorithm is also presented and applied in a robotic platform for object manipulation tasks.

In this paper we investigate an application in which a serial manipulator is engaged in a task driven state transition learning through a set of basic behaviours (i.e. inherited actions). The approach is based on an extension of the SARSA reinforcement learning algorithm. In particular, the case under study consists in the control of the end-effector position sequences of a custom serial manipulator (i.e. the MiniARM) in a constrained shortest path problem. In order to test performances of the overall algorithm and the improvement beyond the state of the art, those strategies have been implemented both in simulation and in a real hardware environment. Results have been analyzed in terms of learning time and iterations needed to complete the assigned task.

The fruit fly Drosophila melanogaster is an extremely interesting insect because it shows a wealth of complex behaviors, despite its small brain. Nowadays genetic techniques allow to knock out the function of defined parts or genes in the Drosophila brain. Together with specific mutants which show similar defects in those parts or genes, hypothesis about the functions of every single brain part can be drawn. Following these experiments, a computational model of the fly Drosophila has been designed with a view to its robotic implementation.

In this paper the implementation of a correlation-based navigation algorithm, based on an unsupervised learning paradigm for spiking neural networks, called Spike Timing Dependent Plasticity (STDP), is presented. The main characteristic of the learning technique implemented is that it allows the robot to learn high-level sensor features, based on a set of basic reflexes, depending on some low-level sensor inputs. The goal is to allow the robot to autonomously learn how to navigate in an unknown environment, avoiding obstacles and heading toward or avoiding the targets (on the basis of the rewarded action). This algorithm was implemented on a bio-inspired hybrid mini-robot, called TriBot. The peculiar characteristic of this robot is its mechanical structure, since it allows to join the advantages both of legs and wheels. In addition, it is equipped with a manipulator that allows to add new capabilities, like carry objects and overcome obstacles. Robot experiments are reported to demonstrate the potentiality and the effectiveness of the approach.

Since many years insects have been considered as a source of inspiration for robotic architectures. From this point of view the fly Drosophila melanogaster is more than likely a protagonist, because of the genetic techniques that allow neurobiologists to make deep studies and hypotheses about the brain of this fly. In this work a computational model of the Drosophila has been tested and implemented on a robot simulator. Moreover, the normal capabilities of the fly have been extended in order to have an useful robot-oriented model. Results about a possible application in a real-life scenario of the whole model of the Drosophila brain are reported.

Insects show the ability to react to certain stimuli with simple reflexes using direct sensory-motor pathways, which can be considered as basic behaviors, while high brain regions provide secondary pathway allowing the emergence of a cognitive behavior which modulates the basic abilities. Taking inspiration from this evidence, a new general purpose perceptual control architecture is briefly presented and experimentally applied to a rover navigating in a cluttered environment. The core of the architecture is constituted by the Representation layer, where different stimuli, triggering competitive reflexes, are fused to form a unique abstract picture of the environment. Each representation induces a learnable modulation of the basic behaviors in order to determine the robot overall behavior. The representation is formalized by means of Reaction-Diffusion nonlinear partial differential equations, under the paradigm of the Cellular Neural Networks (CNNs), whose dynamics converges to steady-state Turing patterns. A suitable unsupervised learning leads to the shaping of the basins of attraction of the Turing patterns that, at the end of the leaning stage, represent a particular behavior modulation. Both simulations and robot experiments are drawn to demonstrate the potentiality and the effectiveness of the approach.

In this paper, we introduce a network of spiking neurons devoted to navigation control. Three different examples, dealing with stimuli of increasing complexity, are investigated. In the first one, obstacle avoidance in a simulated robot is achieved through a network of spiking neurons. In the second example, a second layer is designed aiming to provide the robot with a target approaching system, making it able to move towards visual targets. Finally, a network of spiking neurons for navigation based on visual cues is introduced. In all cases, the robot was assumed to rely on some a priori known responses to low-level sensors (i.e., to contact sensors in the case of obstacles, to proximity target sensors in the case of visual targets, or to the visual target for navigation with visual cues). Based on their knowledge, the robot has to learn the response to high-level stimuli (i.e., range finder sensors or visual input). The biologically plausible paradigm of spike-timing-dependent plasticity (STDP) is included in the network to make the system able to learn high-level responses that guide navigation through a simple unstructured environment. The learning procedure is based on classical conditioning.

In this paper a recently introduced and well promising methodology for robot perception is applied to autonomously learn robot navigation in an unstructured environment. Perception is here considered as the spontaneous, environmentally mediated, emergence of Turing patterns in CNNs as perceptual states. They, plastically associated to suitable actions, lead the robot to solve autonomously its task. Following this concept, robot behavior (in this case navigation) is reflected, in a virtual navigation through the different basins of attraction of the generated patterns, within the robot control neural network. The whole architecture was implemented in an FPGA-based hardware embedded on a roving robot. In the manuscript the perceptual architecture together with experimental results on a roving robot, will be reported.

Correlation based algorithms have been found to explain many basic behaviors in simple animals. In this paper the authors investigate the problem of navigation control of a robot from the viewpoint of bio-inspired perception. In this paper the authors study how to go up, through learning, from the implementation of a reactive system, towards behaviors of increasing complexity. The whole control system is based on networks of spiking neurons. A correlation based rule, namely the spike timing dependent plasticity (STDP), is implemented for an efficient learning. The main interesting consequence is that the system is able to learn high-level sensor features, based on a set of basic reflexes, depending on some low-level sensor inputs. The whole methodology is presented through simulation results and also through its implementation on an FPGA based system for real time working on a roving robot.

This paper describes a general approach for the adaptive supervised learning of behaviors in a behavior-based robot. The key idea is to formalize a behavior produced by a Motor Map driven by an internal adaptive reward function. Aim of the adaptive reward function is to select the most significant sensory inputs and to use them in the best way. The greatest challenge is to keep small the search space. Motor map learning relies on the classical Kohonen algorithm, while the structure of the reward function is learnt through a non-associative reinforcement learning algorithm. Simulation results on a six legged biologically-inspired robot confirm the suitability of the approach. This methodology allows the human designer to easily embody all the a priori knowledge on the robot controller, while providing at the same time a high degree of adaptability and robustness against the sensory malfunctioning

This paper describes the implementation of a bio-inspired six legged robot: Gregor I. Both structure and locomotion control are inspired by biological observations in cockroaches. Robot mechanics attempts to emulate main structural features in cockroaches, like self-stabilizing posture and specializing legged function; in turn, locomotion control is based on the theory of the central pattern generator implemented on a VLSI chip. The final aim is to artificially replicate the fundamental principles that guarantee cockroach's extraordinary agility. Our major concern was on the implementation of rear legs, that seem to play a crucial role in obstacle overcoming and payload capability, and on the locomotion control, performed in this work by a cellular neural network playing the role of an artificial central pattern generator. Experimental tests showed that Gregor I is able to walk at the travel speed of 0.1 body length per second and to successfully negotiate obstacles more than 170 % of the height of its mass center

In the last years, in the area of bio-inspired robotics, the research activity has been directed to high level aspects that includes psychological theories and behavioral approaches. In this paper a new simulation tool for perceptive system based on the sensing-perception-action loop is proposed. The framework has been designed to evaluate the performance of control strategies applied to the navigation of autonomous robots. The tool can be used to create a 3D environment in which the exploring capabilities of a robot executing a navigation task such as for example a food retrieval task, can be evaluated. The behavior of the system is monitored with the help of a 3D real-time visualizer supported by a graphic representations of the trajectory followed

Central Pattern Generators (CPGs) are a suitable paradigm to solve the problem of locomotion control in walking robots. CPGs are able to generate feed-forward signals to achieve a proper coordination among the robot legs. In literature they are often modelled as networks of coupled nonlinear systems. However the topic of feedback in these systems is rarely addressed. On the other hand feedback is essential for locomotion. In this paper the CPG for a hexapod robot is implemented through Cellular Neural Networks (CNNs). Feedback is included in the CPG controller by exploiting the dynamic properties of the CPG motor-neurons, such as synchronization issue and local bifurcations. These universal paradigms provide the essential issues to include sensory feedback in CPG architectures based on coupled nonlinear systems. Experiments on a dynamic model of a hexapod robot are presented to validate the approach introduced.