Dynamic Neural Fields (DNFs) offer a biologically grounded framework for modelling continuous-time neural dynamics underlying a wide range of cognitive and sensorimotor functions. Despite their expressive power, DNF architectures are typically designed and tuned manually, as their nonlinear dynamics and kernel-based interactions make systematic design and parameterisation difficult. As a result, constructing functional DNF models requires expert knowledge and iterative trial-and-error. We introduce NEAT-DNFs, a neuroevolutionary framework that extends NeuroEvolution of Augmenting Topologies (NEAT) to the automated synthesis of DNF architectures. In NEAT-DNFs, neural fields and inter-field interactions are encoded as evolvable genes, enabling the joint evolution of intrinsic parameters and architectural topology. This allows continuous-time neural dynamics to be discovered autonomously, without manual tuning. We evaluate NEAT-DNFs on a hierarchy of benchmark tasks of increasing complexity. Initial tasks target core Dynamic Field Theory (DFT) mechanisms, including detection, working memory, and selection. We then consider more demanding problems that require the emergence of intermediate neural fields, thereby testing the framework's capacity for structural innovation. Across all tasks, NEAT-DNFs reliably evolves compact and stable architectures, increasing structural complexity only when necessary. This work bridges neuroevolution and DFT, establishing a foundation for the automated design of biologically inspired neural field models.

Empirical studies of human mobility patterns reveal a high degree of spatio-temporal regularity, with individuals frequently traveling to the same destinations at relatively the same time. Moreover, the vehicle frequently transports the same occupants carrying specific objects. Indeed, there is a rich body of work on predicting users’ future destinations. However, existing computational approaches primarily rely on offline learning methods that require stationary data distributions, thus not accounting for situations in which traveling habits may change over time. The present research focuses on individual mobility prediction for an advanced driver assistance system within the In-Vehicle Information Systems domain. Specifically, we propose a novel approach based on the framework of Dynamic Neural Fields that learns the traveling routines of a vehicle, including the drivers, passengers, and objects. In our neurodynamics approach, the learning takes place continually and online, allowing the system to accommodate changes in traveling habits over time. The proposed strategy goes beyond learning the mobility routines of a vehicle, enabling predictive assistance for questions like When is the next departure? Where to go? but also extended assistance such as predicting Who is the next driver? Which passengers/objects will enter/leave? We evaluate the performance of the approach in predicting trip information on 11-week datasets of real traveling routines. A vehicle with such cognitive abilities has the potential to improve drivers’ readiness, enhance trip comfort, and prevent objects from being forgotten.

This paper proposes a solution for safe navigation of stacker vehicles in workspaces shared with people, with a focus on the docking manoeuvres for pallet picking and dropping. Behaviours for way-point and wall following are developed following the attractor dynamics approach. Then, these behaviours are orchestrated by state machines (that activate or deactivate them) depending on the specific task. Each of these states also defines different safe areas and maximum travel speeds, which is a requirement for safe operation. Results of real experiments are presented that show the standard operation and its robustness against perturbations (people in the way) and failure detection (missing pallets).

Diagnosis of Fabry disease (FD) remains a challenge mostly due to its rare occurrence and phenotipical variability, with considerable delay between onset and clinical diagnosis. It is then of extreme importance to explore biomarkers capable of assisting the earlier diagnosis of FD. There is growing evidence supporting the use of gait assessment in the diagnosis and management of several neurological diseases. In fact, gait abnormalities have previously been observed in FD, justifying further investigation. The aim of this study is to evaluate the effectiveness of different machine learning strategies when distinguishing patients with FD from healthy controls based on normalized gait features. Gait features of an individual are affected by physical characteristics including age, height, weight, and gender, as well as walking speed or stride length. Therefore, in order to reduce bias due to inter-subject variations a multiple regression (MR) normalization approach for gait data was performed. Four different machine learning strategies - Support Vector Machines (SVM), Random Forest (RF), Multiple Layer Perceptrons (MLPs), and Deep Belief Networks (DBNs) - were employed on raw and normalized gait data. Wearable sensors positioned on both feet were used to acquire the gait data from 36 patients with FD and 34 healthy subjects. Gait normalization using MR revealed significant differences in percentage of stance phase spent in foot flat and pushing ( ), with FD presenting lower percentages in foot flat and higher in pushing. No significant differences were observed before gait normalization. Support Vector Machine was the superior classifier achieving an FD classification accuracy of 78.21% after gait normalization, compared to 71.96% using raw gait data. Gait normalization improved the performance of all classifiers. To the best of our knowledge, this is the first study on gait classification that includes patients with FD, and our results support the use of gait assessment on the clinical assessment of FD.

An automatic algorithm to identify Standard Denavit-Hartenberg parameters of serial manipulators is proposed. The method is based on geometric operations and dual vector algebra to process and determine the relative transformation matrices, from which it is computed the Standard Denavit-Hartenberg (DH) parameters (ai, ai, di, θi). The algorithm was tested in several serial robotic manipulators with varying kinematic structures and joint types: the KUKA LBR iiwa R800, the Rethink Robotics Sawyer, the ABB IRB 140, the Universal Robots UR3, the KINOVA MICO, and the Omron Cobra 650. For all these robotic manipulators, the proposed algorithm was capable of correctly identifying a set of DH parameters. The algorithm source code as well as the test scenarios are publicly available.

We present a motion controller that generates collision free trajectories for autonomous Tugger vehicles operating in dynamic factory environments, where human operators may coexist. The controller is formalized as a dynamic system of path velocity and heading direction, whose vector fields change as sensory information varies. By design the parameters are tuned so that the control variables are close to an attractor of the resultant dynamics most of the time. This contributes to the overall asymptotically stability of the system and makes it robust against perturbations. We present several experiments, in a real factory environment, that highlight different innovative features of the navigation system - flexible and safe solutions for human-aware autonomous navigation in dynamic and cluttered environments. This means, besides generating online collision free trajectories between via points, the system detects the presence of humans, interact with them showing awareness of their presence, and generate adequate motor behavior.

We present a distributed leader-helper architecture for teams of two autonomous mobile robots that jointly transport large payloads while avoiding collisions with obstacles (either static or dynamic). The leader navigates to the goal destination and the helper is responsible for maintaining an appropriate distance (which is a function of the object's length) to the leader. Both robots share the responsibility of ensuring that the transported object does not collide with obstructions. No path needs to be given a priori to the robots nor to the payload. The team is able to perform its transportation task in unknown environments that can have corridors, corners and may change the layout online. The payload can be of different dimensions. The team is able to cope with abrupt/strong perturbations that challenge the team behavior during the execution of the task. These characteristics make this approach suitable to be deployed in warehouses or office-like environments. The motion of each robot is controlled by a time series asymptotically stable states, which is formalized using the attractor dynamics approach to behavior based robotics. The advantages are: (i) the overt behavior is smooth and stable; (ii) because the behavior is generated as a time sequence of attractor states, for the control variables, it contributes to the overall asymptotically stability of the system that makes it robust against perturbations. We present results of experiments in simulated environments and with real robots in real environments.

Many of our everyday tasks require the control of the serial order and the timing of component actions. Using the dynamic neural field (DNF) framework, we address the learning of representations that support the performance of precisely time action sequences. In continuation of previous modeling work and robotics implementations, we ask specifically the question how feedback about executed actions might be used by the learning system to fine tune a joint memory representation of the ordinal and the temporal structure which has been initially acquired by observation. The perceptual memory is represented by a self-stabilized, multi-bump activity pattern of neurons encoding instances of a sensory event (e.g., color, position or pitch) which guides sequence learning. The strength of the population representation of each event is a function of elapsed time since sequence onset. We propose and test in simulations a simple learning rule that detects a mismatch between the expected and realized timing of events and adapts the activation strengths in order to compensate for the movement time needed to achieve the desired effect. The simulation results show that the effector-specific memory representation can be robustly recalled. We discuss the impact of the fast, activation-based learning that the DNF framework provides for robotics applications.

Humans are experts in cooperating in a smooth and proactive manner. Action and intention understanding are critical components of efficient joint action. In the context of the EU Integrated Project JAST [16] we have developed an anthropomorphic robot endowed with these cognitive capacities. This project and respective robot (ARoS) is the focus of the video. More specifically, the results illustrate crucial cognitive capacities for efficient and successful human-robot collaboration such as goal inference, error detection and anticipatory action selection. Results were considered one of the ICT “success stories”[22].

In this paper, we show how a team of autonomous mobile robots, which drive in formation, can be endowed with basic cognitive capabilities. The formation control relies on the leader-follower strategy, with three main pair-wise configurations: column, line and oblique. Furthermore, non-linear attractor dynamics are used to generate basic robotic behaviors (i.e. follow-the-leader and avoid obstacles). The control architecture of each follower integrates a representation of the leader (target) direction, which supports leader detection, selection between multiple leaders (decision) and temporary estimation of leader direction (short-term memory during occlusion and prediction). Formalized as a dynamic neural field, this additional layer is smoothly integrated with the motor movement control system. Experiments conducted in our 3D simulation software, as well as results from the implementation in middle size robotic platforms, show the ability for the team to navigate, whilst keeping formation, through unknown and unstructured environments and is robust against ambiguous and temporarily absent sensory information.

This talk presents recent functional insights derived from behavioural and neuroimaging studies into the cognitive mechanisms underlying human joint action. The main question is how the cognitive system of one actor can organize the perceptual consequences of the movements of another actor such that effective joint action in the two actors can take place. Particularly, the issue of complementing each other's action in contrast to merely imitating the actions that one observes will be discussed. Other issues that have been investigated are motivational states (cooperative or competitive), error-monitoring and the interaction between actors at the level of motor control. The talk is completed with presenting recent attempts to implement the functional insights from these experiments into autonomous systems being capable of collaborating intelligently on shared motor tasks.

Dynamical systems theory is used as a theoretical language and tool to design a distributed control architecture for teams of mobile robots, that must transport a large object and simultaneously avoid collisions with (either static or dynamic) obstacles. Here we demonstrate in simulations and implementations in real robots that it is possible to simplify the architectures presented in previous work and to extend the approach to teams of n robots. The robots have no prior knowledge of the environment. The motion of each robot is controlled by a time series of asymptotical stable states. The attractor dynamics permits the integration of information from various sources in a graded manner. As a result, the robots show a strikingly smooth an stable team behaviour.

This paper describes the use of non-linear attractor dynamics to generate trajectories free of collisions for an autonomous vehicle moving in 3D. Computer simulations are used to test the behaviour and stability of the airship motion during the task of flying toward a target destination while simultaneously avoiding collisions with (static or moving) obstacles. Results indicate that if parameter values are chosen within reasonable ranges, then the flight controller works quite well even when subject to disturbances along the generated flight path.