Insect-inspired navigation strategies have the potential to unlock robotic navigation in power-constrained scenarios, as they can function effectively with limited computational resources. One such strategy, familiarity-based navigation, has successfully navigated a robot along routes of up to 60 m using a single-layer neural network trained with an Infomax learning rule. Given the small size of the network that effectively encodes the route, here we investigate the limits of this method, challenging it to navigate longer routes, investigating the relationship between performance, view acquisition rate and dimension, network size, and robustness to noise. Our goal is both to determine the parameters at which this method operates effectively and to explore the profile with which it fails, both to inform theories of insect navigation and to improve robotic deployments. We show that effective memorization of familiar views is possible for longer routes than previously attempted, but that this length decreases for reduced input view dimensions. We also show that the ideal view acquisition rate must be increased with route length for consistent performance. We further demonstrate that computational and memory savings may be made with equivalent performance by reducing the network size-an important consideration for applicability to small, lower-power robots-and investigate the profile of memory failure, demonstrating increased confusion across the route as it extends in length. In this extension to previous work, we also investigate the form taken by the network weights as training extends and the areas of the image on which visual familiarity-based navigation most relies. Additionally, we investigate the robustness of familiarity-based navigation to view variation caused by noise.

Individually foraging ants use egocentric views as a dominant navigation strategy for learning and retracing routes. Evidence suggests that route retracing can be achieved by algorithms which use views as 'visual compasses', where individuals choose the heading that leads to the most familiar visual scene when compared to route memories. However, such a mechanism does not naturally lead to route approach, and alternative strategies are required to enable convergence when off-route and for correcting on-route divergence. In this work we investigate how behaviour incorporated into visual compass like route learning and recapitulation strategies might enable convergence to a learned route and its destination. Without alterations to the basic form of the initial learning route, the most successful recapitulation method comes from a 'cast and surge' approach, a mechanism seen across arthropods for olfactory navigation. In this strategy casts form a 'zig-zagged' or oscillatory search in space for familiar views, and surges exploit visual familiarity gradients. We also find that performance improves if the learned route consists of an oscillatory motor mechanism with learning gated to occur when the agent approaches the central axis of the oscillation. Furthermore, such oscillations combined with the cast and surge method additively enhance performance, showing that it benefits to incorporate oscillatory behaviour in both learning and recapitulation. As destination reaching is the primary goal of navigation, we show that a suitably sized goal-orientated learning walk might suffice, but that the scale of this is dependent on the degree of divergence, and thus depends on route length and the route learning and recapitulation strategies used. Finally we show that view familiarity can modulate on-the-spot scans performed by an agent, providing a better reflection of ant behaviour. Overall, our results show that the visual compass can provide a basis for robust visual navigation, so long as it is considered holistically with the details of basic motor and sensory-motor patterns of ants undertaking route learning and recapitulation.

Many insects use memories of their visual environment to adaptively drive spatial behaviours. In ants, visual memories are fundamental for navigation, whereby foragers follow long visually guided routes to foraging sites and return to the location of their nest. Whilst we understand the basic visual pathway to the memory centres (Optic Lobes to Mushroom Bodies) involved in the storage of visual information, it is still largely unknown what type of representation of visual scenes underpins view-based navigation in ants. Several experimental studies have suggested ants use "higher-order" visual information - that is features extracted across the whole extent of a visual scene - which raises the question as to how these features might be computed. One such experimental study showed that ants can use the proportion of a shape experienced left of their visual centre to learn and recapitulate a route, a feature referred to as "fractional position of mass" (FPM). In this work, we use a simple model constrained by the known neuroanatomy and information processing properties of the Mushroom Bodies to explore whether the apparent use of the FPM could be a resulting factor of the bilateral organisation of the insect brain, all the whilst assuming a simple "retinotopic" view representation. We demonstrate that such bilaterally organised memory models can implicitly encode the FPM learned during training. We find that balancing the "quality" of the memory match across left and right hemispheres allows a trained model to retrieve the FPM defined direction, even when the model is tested with novel shapes, as demonstrated by ants. The result is shown to be largely independent of model parameter values, therefore suggesting that some aspects of higher-order processing of a visual scene may be emergent from the structure of the neural circuits, rather than computed in discrete processing modules.

The central complex of insects contains cells, organised as a ring attractor, that encode head direction. The 'bump' of activity in the ring can be updated by idiothetic cues and external sensory information. Plasticity at the synapses between these cells and the ring neurons, that are responsible for bringing sensory information into the central complex, has been proposed to form a mapping between visual cues and the heading estimate which allows for more accurate tracking of the current heading, than if only idiothetic information were used. In Drosophila, ring neurons have well characterised non-linear receptive fields. In this work we produce synthetic versions of these visual receptive fields using a combination of excitatory inputs and mutual inhibition between ring neurons. We use these receptive fields to bring visual information into a spiking neural network model of the insect central complex based on the recently published Drosophila connectome. Previous modelling work has focused on how this circuit functions as a ring attractor using the same type of simple visual cues commonly used experimentally. While we initially test the model on these simple stimuli, we then go on to apply the model to complex natural scenes containing multiple conflicting cues. We show that this simple visual filtering provided by the ring neurons is sufficient to form a mapping between heading and visual features and maintain the heading estimate in the absence of angular velocity input. The network is successful at tracking heading even when presented with videos of natural scenes containing conflicting information from environmental changes and translation of the camera.

Reliably identifying water bodies is an important step in automating the identification of potable water. This work therefore investigates water scene segmentation, with a focus on water’s physical properties which give it features that distinguish it from other elements. We propose a physics-aware water segmentation method, in which we adapt both a U-Net and MACUNet model so that they are biased towards texture information by using a Gabor convolutional layer as the first layer, combined with a mixture of average and maximum pooling layers in the encoder. To train the networks, a dataset of annotated water images was created, comprising water bodies from diverse light, atmospheric, geographic, and environmental conditions. We show that the physics-aware, texture-biased models result in effective water segmentation. We then test the texture-biased models using 3 standard aerial scene segmentation benchmarks and show that in all cases they outperform the standard U-Net or MACUNet models. We suggest this is because the new models are sensitive to small variations in texture, meaning they can extract information from scenes affected by light, canopy or shadows.

In certain geographic locations like South Asia, the landscape changes dramatically between dry and wet seasons. The main factor responsible for this variation is the flora that trans-forms the landscape between seasons. These transformations can affect the performance of deep learning models trained to analyse satellite images, especially if there are domain shifts between training and testing data distributions. The current work shows that an architecture which employs a Gabor convolutional layer as the first layer of a deep network input fo-cuses on more salient parts of the image than one which uses a standard convolutional layer meaning that removing colour information is less damaging than for the standard network. Further we show that the proposed architecture is robust in the presence of domain shifts due to seasonal data variations.

Abstract This paper explores current developments in evolutionary and bio-inspired approaches to autonomous robotics, concentrating on research from our group at the University of Sussex. These developments are discussed in the context of advances in the wider fields of adaptive and evolutionary approaches to AI and robotics, focusing on the exploitation of embodied dynamics to create behaviour. Four case studies highlight various aspects of such exploitation. The first exploits the dynamical properties of a physical electronic substrate, demonstrating for the first time how component-level analog electronic circuits can be evolved directly in hardware to act as robot controllers. The second develops novel, effective and highly parsimonious navigation methods inspired by the way insects exploit the embodied dynamics of innate behaviours. Combining biological experiments with robotic modeling, it is shown how rapid route learning can be achieved with the aid of navigation-specific visual information that is provided and exploited by the innate behaviours. The third study focuses on the exploitation of neuromechanical chaos in the generation of robust motor behaviours. It is demonstrated how chaotic dynamics can be exploited to power a goal-driven search for desired motor behaviours in embodied systems using a particular control architecture based around neural oscillators. The dynamics are shown to be chaotic at all levels in the system, from the neural to the embodied mechanical. The final study explores the exploitation of the dynamics of brain-body-environment interactions for efficient, agile flapping winged flight. It is shown how a multi-objective evolutionary algorithm can be used to evolved dynamical neural controllers for a simulated flapping wing robot with feathered wings. Results demonstrate robust, stable, agile flight is achieved in the face of random wind gusts by exploiting complex asymmetric dynamics partly enabled by continually changing wing and tail morphologies.

For the first time, a field programmable transistor array (FPTA) was used to evolve robot control circuits directly in analog hardware. Controllers were successfully incrementally evolved for a physical robot engaged in a series of visually guided behaviours, including finding a target in a complex environment where the goal was hidden from most locations. Circuits for recognising spoken commands were also evolved and these were used in conjunction with the controllers to enable voice control of the robot, triggering behavioural switching. Poor quality visual sensors were deliberately used to test the ability of evolved analog circuits to deal with noisy uncertain data in realtime. Visual features were coevolved with the controllers to automatically achieve dimensionality reduction and feature extraction and selection in an integrated way. An efficient new method was developed for simulating the robot in its visual environment. This allowed controllers to be evaluated in a simulation connected to the FPTA. The controllers then transferred seamlessly to the real world. The circuit replication issue was also addressed in experiments where circuits were evolved to be able to function correctly in multiple areas of the FPTA. A methodology was developed to analyse the evolved circuits which provided insights into their operation. Comparative experiments demonstrated the superior evolvability of the transistor array medium.

The Jacobian matrix is a highly popular tool for the control and performance analysis of closed-loop robots. Its usefulness in parallel mechanisms is certainly apparent, and its application to solve motion planning problems, or other higher level questions, has been seldom queried, or limited to nonredundant systems. In this article, we discuss the shortcomings of the use of the Jacobian matrix under redundancy, in particular when applied to kinematically redundant parallel architectures with non-serially connected actuators. These architectures have become fairly popular recently as they allow the end-effector to achieve full rotations, which is an impossible task with traditional topologies.The problems with the Jacobian matrix in these novel systems arise from the need to eliminate redundant variables forming it, resulting in both situations where the Jacobian incorrectly identifies singularities (false positive), and where it fails to identify singularities (false negative). These issues have, thus far, remained unaddressed in the literature. We highlight these limitations herein by demonstrating several cases using numerical examples of both planar and spatial architectures.

All organisms wishing to survive and reproduce must be able to respond adaptively to a complex, changing world. Yet the computational power available is constrained by biology and evolution, favouring mechanisms that are parsimonious yet robust. Here we investigate the information carried in small populations of visually responsive neurons in Drosophila melanogaster. These so-called 'ring neurons', projecting to the ellipsoid body of the central complex, are reported to be necessary for complex visual tasks such as pattern recognition and visual navigation. Recently the receptive fields of these neurons have been mapped, allowing us to investigate how well they can support such behaviours. For instance, in a simulation of classic pattern discrimination experiments, we show that the pattern of output from the ring neurons matches observed fly behaviour. However, performance of the neurons (as with flies) is not perfect and can be easily improved with the addition of extra neurons, suggesting the neurons' receptive fields are not optimised for recognising abstract shapes, a conclusion which casts doubt on cognitive explanations of fly behaviour in pattern recognition assays. Using artificial neural networks, we then assess how easy it is to decode more general information about stimulus shape from the ring neuron population codes. We show that these neurons are well suited for encoding information about size, position and orientation, which are more relevant behavioural parameters for a fly than abstract pattern properties. This leads us to suggest that in order to understand the properties of neural systems, one must consider how perceptual circuits put information at the service of behaviour.

Compliant bodies with complex dynamics can be used both to simplify control problems and to lead to adaptive reflexive behavior when engaged with the environment in the sensorimotor loop. By revisiting an experiment introduced by Beer and replacing the continuous-time recurrent neural network therein with reservoir computing networks abstracted from compliant bodies, we demonstrate that adaptive behavior can be produced by an agent in which the body is the main computational locus. We show that bodies with complex dynamics are capable of integrating, storing, and processing information in meaningful and useful ways, and furthermore that with the addition of the simplest of nervous systems such bodies can generate behavior that could equally be described as reflexive or minimally cognitive.

We propose a biologically plausible architecture for unsupervised ensemble learning in a population of spiking neural network classifiers. A mixture of experts type organisation is shown to be effective, with the individual classifier outputs combined via a gating network whose operation is driven by input timing dependent plasticity (ITDP). The ITDP gating mechanism is based on recent experimental findings. An abstract, analytically tractable model of the ITDP driven ensemble architecture is derived from a logical model based on the probabilities of neural firing events. A detailed analysis of this model provides insights that allow it to be extended into a full, biologically plausible, computational implementation of the architecture which is demonstrated on a visual classification task. The extended model makes use of a style of spiking network, first introduced as a model of cortical microcircuits, that is capable of Bayesian inference, effectively performing expectation maximization. The unsupervised ensemble learning mechanism, based around such spiking expectation maximization (SEM) networks whose combined outputs are mediated by ITDP, is shown to perform the visual classification task well and to generalize to unseen data. The combined ensemble performance is significantly better than that of the individual classifiers, validating the ensemble architecture and learning mechanisms. The properties of the full model are analysed in the light of extensive experiments with the classification task, including an investigation into the influence of different input feature selection schemes and a comparison with a hierarchical STDP based ensemble architecture.

We present the first example of ‘minimally cognitive’ sensorimotor behaviour arising from a body as physical reservoir. \nBy revisiting an experiment introduced by Beer (1996) \nand replacing the continuous-time recurrent neural network \n(CTRNN) therein with networks of mass-spring-dampers we \ndemonstrate that bodies may be exploited for more than control \nand pattern generation and take over some tasks which \nwere previously thought to require a central nervous system.

When niching or speciation is required to perform a task that has several different component parts, standard genetic algorithms (GAs) struggle. They tend to evaluate and select all individuals on the same part of the task, which leads to genetic convergence within the population. The goal of evolutionary niching methods is to enforce diversity in the population so that this genetic convergence is avoided. One drawback with some of these niching methods is that they require a priori knowledge or assumptions about the specific fitness landscape in order to work; another is that many such methods are not set up to work on cooperative tasks where fitness is only relevant at the group level. Here we address these problems by presenting the group GA, described earlier by the authors, which is a group-based evolutionary algorithm that can lead to emergent niching. After demonstrating the group GA on an immune system matching task, we extend the previous work and present two modified versions where the number of niches does not need to be specified ahead of time. In the random-group-size GA, the number of niches is varied randomly during evolution, and in the evolved-group-size GA the number of niches is optimized by evolution. This provides a framework in which we can evolve groups of individuals to collectively perform tasks with minimal a priori knowledge of how many subtasks there are or how they should be shared out.

In this paper a new, evolutionary multi-objective approach is introduced to tool sequence optimization in rough milling. Previous research has focused on the optimization of either the tool sequence or associated cutting parameters. Here, the tool sequence and a machining parameter, the cutting speeds of the individual tools, are simultaneously optimized, producing a Pareto front with both discrete and continuous properties. This is the first time that a multiple-tool multi-objective approach has been taken to tool selection, offering a set of solutions to the process planner. Three objectives are considered, thickness of excess stock, machining time and tooling costs. Unconstrained NSGA-II is used as the base algorithm but several preferential search strategies are tested to attempt to deal with constraints and guide search towards the Pareto optimal front. These include the established reference point (R-NSGA-ii) and weighted objective (WO) methods, as well as two novel techniques - “Guided Elitism” (GE) and “Precedential Objective Order Ranking” (PR). While WO performs best on average when assessed using the hypervolume indicator, the algorithms behave differently in terms of the quality and diversity of solutions found. A hybrid method using GE for exploration and PR for exploitation is shown to outperform the other techniques across all performance measures.

Selecting the sequence of tools to use for the rough machining of components is an important task in manufacturing, which greatly affects the overall machining time and cost of the process. In this paper a multi-objective approach is presented, which supports the use of tools with different geometrical properties and offers the process planner a set of Pareto optimal solutions. An industrial simulator is employed, which allows important information to be captured in the model but has the disadvantage of being computationally expensive. A master/slave approach to parallelization is implemented, which can be used on existing grid or cloud computing infrastructures. Synchronous generational and asynchronous steady-state multi-objective algorithms are compared on their search performance and runtimes on two components. Particular attention is paid to potential problems faced by asynchronous search caused by heterogeneous evaluation times due to characteristics present in individual tool sequences. Results show that the algorithms achieve a similar search performance, with the synchronous algorithm occasionally finding a slightly more diverse spread of solutions. However, the asynchronous algorithm is considerably faster, and provides good solutions in a short runtime that means this approach could be easily and inexpensively implemented in an industrial setting.

In this paper we discuss our approach to solving the tool selection problem, specifically applied to rough machining. A simulation is used to evaluate tool sequences, which provides accurate values for tool paths and a 3D model of the final machined part. This allows for a largely unrestricted search using different tool types, making this approach more useful for real world applications than previous attempts at solving the problem. An exhaustive search of every valid tool sequence is executed and shows that assumptions present in related research can prevent the optimal solution from being discovered. Metaheuristic algorithms are used to traverse the search space because of its complex combinatorial properties. Four algorithms are tested - Genetic Algorithm, Stochastic Hill Climbing, Hybrid Genetic Algorithm and Random Restart Stochastic Hill Climbing. Evaluating their performance at coping with two competing demands, finding optimal solutions and keeping the number of potentially expensive evaluations low, it is shown that RRSHC performs best in terms of solution accuracy but at the greatest computational cost. SHC finds the optimum sequence less frequently but needs far fewer evaluations and the HGA lies somewhere in between, making it a good choice if the problem domain is not well-specified.

In this paper we propose a model of visually guided route navigation in ants that captures the known properties of real behaviour whilst retaining mechanistic simplicity and thus biological plausibility. For an ant, the coupling of movement and viewing direction means that a familiar view specifies a familiar direction of movement. Since the views experienced along a habitual route will be more familiar, route navigation can be re-cast as a search for familiar views. This search can be performed with a simple scanning routine, a behaviour that ants have been observed to perform. We test this proposed route navigation strategy in simulation, by learning a series of routes through visually cluttered environments consisting of objects that are only distinguishable as silhouettes against the sky. In the first instance we determine view familiarity by exhaustive comparison with the set of views experienced during training. In further experiments we train an artificial neural network to perform familiarity discrimination using the training views. Our results indicate that, not only is the approach successful, but also that the routes that are learnt show many of the characteristics of the routes of desert ants. As such, we believe the model represents the only detailed and complete model of insect route guidance to date. What is more, the model provides a general demonstration that visually guided routes can be produced with parsimonious mechanisms that do not specify when or what to learn, nor separate routes into sequences of waypoints.

Rate-coded Hebbian learning, as characterized by the BCM formulation, is an established computational model of synaptic plasticity. Recently it has been demonstrated that changes in the strength of synapses in vivo can also depend explicitly on the relative timing of pre- and postsynaptic firing. Computational modeling of this spike-timing-dependent plasticity (STDP) has demonstrated that it can provide inherent stability or competition based on local synaptic variables. However, it has also been demonstrated that these properties rely on synaptic weights being either depressed or unchanged by an increase in mean stochastic firing rates, which directly contradicts empirical data. Several analytical studies have addressed this apparent dichotomy and identified conditions under which distinct and disparate STDP rules can be reconciled with rate-coded Hebbian learning. The aim of this research is to verify, unify, and expand on these previous findings by manipulating each element of a standard computational STDP model in turn. This allows us to identify the conditions under which this plasticity rule can replicate experimental data obtained using both rate and temporal stimulation protocols in a spiking recurrent neural network. Our results describe how the relative scale of mean synaptic weights and their dependence on stochastic pre- or postsynaptic firing rates can be manipulated by adjusting the exact profile of the asymmetric learning window and temporal restrictions on spike pair interactions respectively. These findings imply that previously disparate models of rate-coded autoassociative learning and temporally coded heteroassociative learning, mediated by symmetric and asymmetric connections respectively, can be implemented in a single network using a single plasticity rule. However, we also demonstrate that forms of STDP that can be reconciled with rate-coded Hebbian learning do not generate inherent synaptic competition, and thus some additional mechanism is required to guarantee long-term input-output selectivity.

The firing rate of single neurons in the mammalian hippocampus has been demonstrated to encode for a range of spatial and non-spatial stimuli. It has also been demonstrated that phase of firing, with respect to the theta oscillation that dominates the hippocampal EEG during stereotype learning behaviour, correlates with an animal's spatial location. These findings have led to the hypothesis that the hippocampus operates using a dual (rate and temporal) coding system. To investigate the phenomenon of dual coding in the hippocampus, we examine a spiking recurrent network model with theta coded neural dynamics and an STDP rule that mediates rate-coded Hebbian learning when pre- and post-synaptic firing is stochastic. We demonstrate that this plasticity rule can generate both symmetric and asymmetric connections between neurons that fire at concurrent or successive theta phase, respectively, and subsequently produce both pattern completion and sequence prediction from partial cues. This unifies previously disparate auto- and hetero-associative network models of hippocampal function and provides them with a firmer basis in modern neurobiology. Furthermore, the encoding and reactivation of activity in mutually exciting Hebbian cell assemblies demonstrated here is believed to represent a fundamental mechanism of cognitive processing in the brain.

No abstract available.

Recent years have seen the discovery of freely diffusing gaseous neurotransmitters, such as nitric oxide (NO), in biological nervous systems. A type of artificial neural network (ANN) inspired by such gaseous signaling, the GasNet, has previously been shown to be more evolvable than traditional ANNs when used as an artificial nervous system in an evolutionary robotics setting, where evolvability means consistent speed to very good solutions--here, appropriate sensorimotor behavior-generating systems. We present two new versions of the GasNet, which take further inspiration from the properties of neuronal gaseous signaling. The plexus model is inspired by the extraordinary NO-producing cortical plexus structure of neural fibers and the properties of the diffusing NO signal it generates. The receptor model is inspired by the mediating action of eurotransmitter receptors. Both models are shown to significantly further improve evolvability. We describe a series of analyses suggesting that the reasons for the increase in evolvability are related to the flexible loose coupling of distinct signaling mechanisms, one "chemical" and one "electrical."

The twin fitness landscape properties of neutrality and ruggedness are crucial to the dynamics of evolutionary optimisation. In this paper, we investigate the interplay between these two properties in a complex evolutionary robotics fitness landscape, through the introduction of four robot controller architecture models; the GasNet, uniform, dispersed and plexus models. We show that, in isolation, neither added neutrality or decreased ruggedness (coupling) in the models produces increase in the speed of evolution. However, both effects in conjunction produce a significant increase in the speed of evolution.