BIOTACT.org

Research in the BIOTACT project is split into four key activities. Please click the links for details on each.

Artificial vibrissal sensing is a relatively new field. Basic questions concerning the ideal shape, and material properties of artificial whiskers; how best to transduce whisker deflections into sensor signals; and how to actuate the whiskers, either individually or as an array, remain to be answered. We will explore a variety of alternative solutions to those previously tested in order to get closer, in terms of functional properties, accuracy, and size, to the biological models we are trying to emulate. This work will be supported by biological investigations of the whisker, follicle, and early sensory processing stations, and by computational modelling of these components.

Specific issues to be addressed include:

• Material properties of the whisker shaft
• Sensor transduction
• Sensor actuation
• Sensor design
• Early sensory processing
• Technological advancement

Material properties of the whisker shaft

Facial whiskers are delicate, tapered elastic beams that deform easily upon touch, and probably also during whisking in air. Mechanical deformations convey specific sensory information about air currents and external objects via morphological effects such as changes in whisker curvature. The reliability and information content of this kind of “morphological computation”, and its use during active touch, will be investigated in order to optimise the design of this component.

Sensor transduction

Mechanoreceptors in the whisker follicle respond with high acuity to deflections of the whisker. Although several useful prototypes have been developed, the optimal design of transduction mechanisms for artificial whiskers has yet to be determined. Both existing prototypes and novel designs will be tested and their performance compared to that of their biological counterparts.

Sensor actuation

Muscle-like micro-actuators are currently a highly active research field in engineering. The challenge will be to design an actuated sensing device that is reliable, energy efficient, and of sufficiently small in size, but also provides fast and accurate actuation with adequate degrees of freedom of movement to replicate much of the versatility of natural whisking control.

Sensor design

The structure of the BIOTACT sensor will be inspired by the organisation of the rat vibrissae, where the longer, more widely-spaced, actuated macrovibrissae surround a dense array of shorter, non-actuated microvibrissae. Different technological solutions will likely be adopted for these two components. It is anticipated that the final device will be standalone, in the sense that it could be attached to a variety of different robot platforms. Indeed, for the purposes of development and testing, prototypes will mounted as the end-effector of a robot manipulator allowing them to be manoeuvred into any position in 3d space.

Early sensory processing

Each whisker follicle in the rat is innervated by up to 200 sensory fibres. Thus ascending signals are processed first in the trigeminal ganglion, then the trigeminal complex, and via loops involving the cerebellum, before being relayed to further processing stations in the mid- and fore- brain. Building on existing pilot work by consortium partners we will develop biomimetic models of these early sensory processing stations, to determine their effects on relayed signals. For example, sensory information from whisking contains artefacts produced by the animal's own movements. We will therefore employ a model of the adaptive cerebellar loop to test the hypothesis that the cerebellum monitors motor control signals and proprioceptive signals and uses these to remove self-movement artefacts from whisking information. Special purpose hardware in the form of application-specific integrated circuits (ASICs) will also be developed to provide compact, real-time implementation of these signal-processing operations.

Technological enhancements

We believe there are very many useful insights to be mined from the study of natural vibrissal touch systems. However, slavishly copying biology does not necessarily generate the most effective technologies. Therefore, in addition to developing a strongly biomimetic version of the BIOTACT sensor for use as a research tool, in the final year of the project we will look for ways to optimise and streamline the design without further specific regard for biological accuracy. We will then explore the potential of this technologically-enhanced sensor for applications in industrial settings.

The fast, graceful, coordinated movements made by animals and humans are very different from the clumsy stereotyped movements that we characterise as “robotic”; neural control algorithms employed by biological systems probably contribute significantly to this superiority. In this activity, we will investigate and simulate the neural circuitry underlying sensor actuation and control in the vibrissal system. Our objective will be to understand how this system operates in different task settings, and how it is modulated by incoming sensory information so as to generate the intelligent, information-seeking behaviour we see in natural active touch. As part of this work, detailed investigations will be made of whisking in freely-moving animals using state-of-the-art technologies such as high-speed video recording and tracking, and wireless electrophysiology. We believe that results from these studies will be vital for devising artificial vibrissal technologies since, based on our active sensing hypothesis, we consider the control strategies used to position the sensory apparatus is as important an aspect of sensing as the properties of the sensor and of the stimulus. Sensorimotor control strategies observed in animals will be compared with, and where appropriate, modelled using techniques from control theory in which a number of project partners have significant track record. The following inter-twined issues will be investigated in depth:

• The nested-loop architecture for sensorimotor control
• Rhythmogenesis
• Task dependency in active touch
• The modulation of whisking pattern generation

The nested-loop architecture for sensorimotor control

A fundamental goal will be to explore the advantages for active sensing provided by the multiple sensorimotor loops in the neural control architecture for vibrissal touch. Large-scale closed loops have interesting computational properties that may optimise sensory processing. In the vibrissal system, these loops are organized in what appears to be a hierarchy, which likely preserves an evolutionary, and perhaps also a developmental, order. Higher levels generally add increased competence by modulating the behaviour/activity of lower levels. Research in robot control has demonstrated that layered control systems of this sort can combine the advantages of fast responsiveness with resistance to damage. Understanding the detailed operation of this complex architecture for vibrissal control should therefore generate insights that will aid the design of effective and robust controllers for artefacts such as autonomous robots.

Rhythmogenesis

Movements of the whiskers are driven from a population of “vibrissal motor neurons” (VMNs) in the brainstem facial nucleus whose rhythmic firing generates the rapid sweeping behaviour termed “whisking”. However, many other areas of the vibrissal sensorimotor system also exhibit some intrinsic rhythmic capacity. Understanding how whisking rhythms are generated and become synchronized (or desynchronized) through the interactions of different cell populations will provide a rich source of general insight into pattern formation in neural systems.

Task dependency in active touch

Human tactile sensing makes use of different strategies for positioning the sensory apparatus depending upon the task in hand. In fact, previous human studies have demonstrated that selection of non-optimal motor strategies can yield significant perceptual impairments that can be corrected by re-learning the task using optimal ones. Anecdotal evidence suggests that control of sensor movement in vibrissal sensory systems likewise adapts according to the current motivation of the animal. By exploring and analysing this task-dependency in detail, we will both uncover useful control strategies for active touch, and further our understanding of how motor control and signal processing in biological sensory systems vary with task demands.

The modulation of whisking pattern generation

Observations of whisking in animals that are stationary and whisking in air (‘free whisking’) reliably demonstrate bilateral symmetry and synchrony of whisker movements on the two sides of the snout. Likewise, within the two whisker fields, the whiskers move largely in phase with one another and at similar velocities. This had led to the assumption that whisking is a relatively stereotyped action that can be reasonably well characterized by describing the amplitude, frequency, and set-point of the overall whisking pattern. However, recent evidence shows that the whisking movements of freely-moving animals engaged in exploratory behaviour vary considerably and consistently in anticipation of head movements, or as a result of contacts with nearby surfaces. Evidence of significant variation in whisker movements within each whisker field is also beginning to accumulate. Investigation of how proprioceptive and sensory data modulates vibrissal control should provide insights applicable to understanding active sensing strategies in other species and modalities.

In this activity we will develop novel computational methods for the extraction of tactile properties of detected objects (texture, shape, etc.). We will also develop algorithms for classification, decision-making, and spatial mapping, based on these tactile codes. Work will focus on the following functional capacities and related computational principles:

• Tactile coding of object properties
• Classification, memory & decision-making

Tactile coding of object properties

Whiskered animals have remarkable capacities for detecting and responding to salient properties of surfaces and objects in their environment. For instance, rats are excellent judges of distance and relative position, and can perform fast and accurate discriminations based on shape and texture—often surpassing the performance of human subjects on similar tasks. Etruscan shrews are the world’s smallest mammal with a brain-size 20,000 times smaller than the human brain. Nevertheless, these animals also show remarkable tactile object recognition capacities indicating that high-level abstract sensory representations are neither restricted to the primates nor tied to large brain size. We will gather observations of the whisker-mediated behavioural capacities of rats and shrews interacting with objects both during free exploration and in tactile discrimination tasks. We will also record neuronal activity from multiple stations of the sensory pathway, from periphery to cortex, during these behaviours. Since the normal functioning of the sensory system entails the simultaneous solving of both spatial problems (where is the object?) and identification problems (what is the object?), we will pay close attention to the interaction of the two sorts of information. The goal of this activity will be to establish how information about tactile object properties can be effectively coded, and how such codes might be modified according to the behaviour in progress. Modelling approaches will include spiking neuron models, Bayesian models of pattern recognition, and abstract functional models of cortical microcircuits.

Classification, memory, and decision-making

Besides the accuracy of tactile behaviours in whiskered animals, a remarkable characteristic is the speed of transformation of complex sensory signals into decisions and the rapid execution of the appropriate choice action. Timescales, from start to finish, can be of the order of tens of milliseconds. In contrast, in the visual domain even task-experienced human observers take 300 ms to indicate the detection of a simple visual object. The rapidity of vibrissal-guided control implies a quick and accurate transformation of raw sense signals into activity states that represent distinct classes of task-relevant stimuli; each class associated with and transformed to its appropriate action. Understanding how biological systems accomplish this, in noisy and uncertain environments, will be critical to endowing biomimetic systems with accuracy, speed, and flexibility. Neurobiological investigations of this question will focus on neural circuits involving the sensory cortex, the hippocampus, and the prefrontal cortex, where we will seek to uncover the representation of information as it resides in short-term memory “buffers”. Experiments involving electrophysiological recording during behaviour will be complemented by others using stimulation of single neurons to assess the behavioural impact of their induced activity. Modelling approaches will use recurrent neural networks and biologically-realistic Hebbian learning rules, together with statistical models of optimal decision-making originally developed using human and primate data. The efficacy of the resulting systems will be tested using the BIOTACT sensor in tasks that involve making tactile discriminations between objects with a speed and accuracy intended to approach that seen in whiskered animals.

One of the most interesting and demanding uses of the vibrissal sense in both rodents and insectivores is in predation. Etruscan shrews, for instance, prey on insects such as crickets which are themselves highly agile, exquisitely mechanosensitive, and almost as large as the shrew itself. Shrews succeed in hunting these creatures by fast and precise attacks. Previous work has demonstrated that tactile shape cues are both necessary and sufficient for evoking and controlling these attacks, whereas visual and olfactory cues are not needed. Rats are also efficient predators that can detect, track, and immobilise prey animals such as cockroaches in darkness. In this activity we will reverse-engineer the neurobiological substrates for vibrissal tracking to develop algorithms that will allow a wheeled mobile robot, equipped with an artificial active vibrissal system and a highly manoeuvrable “head”, to locate, identify, and track fast-moving targets such as smaller robots or remote-controlled toy vehicles. This work will require scientific and technological advances in the following areas:

• Biomimetic control of orienting
• Biomimetic control in tracking
• Fusing predictive tracking with object recognition

Biomimetic control of orienting

Successful predation in whiskered animals is reliant on the midbrain superior colliculus which is involved in controlling the rapid orienting movements needed to locate and keep pace with a moving target. In addition to vibrissal information the colliculus will also require vestibular and/or proprioceptive inputs in order to generate appropriate control signals. We will perform electrophysiological and behavioural studies to investigate the role of the colliculus in the control of whisking and whisker-guided head movements, and to understand the integration of vestibular and whisker information in animals that are searching for a target in space. Computational models will be developed to explain these data and to control whisking/orienting movements in simulated and actual whisking robots.

Biomimetic predictive control in tracking

The fast, ballistic movements needed for prey tracking will also require predictive rather reactive control. The cerebellum is likely to play an essential role here as cerebellar damage particularly impairs predictive aspects of motor behaviour. Studies of the role of the cerebellum in the control of primate grasp also suggest it as a likely neural substrate for relevant internal, forward models. Anatomically, the cerebellum is involved in a closed loop with the superior colliculus indicating a close synergy between these two systems. Building on models of cerebellar predictive control developed for other motor behaviours, such as eye movements, we will develop cerebellar-inspired algorithms for predictive pursuit and evaluate their effectiveness in controlling tactile tracking behaviour in our robot platform.

Fusing predictive tracking with object recognition

To complete our model of predation in whisking animals, and to generate a biomimetic system that demonstrates useful integrative functionality, we will combine the orienting/tracking competences described above with those for tactile object recognition investigated in activity 3 (Adaptive tactile coding, memory & decision-making). The resulting control systems will allow our whiskered mobile robot platform to distinguish between target objects to be tracked and distracter objects to be ignored or avoided. We believe that the demonstration of such functionality will represent a very substantial advance in artificial tactile sensing and control compared to existing technologies.