A link to the seminars given last year (2003-2004) is provided so that the program is still available.
Matthew Szenher, Visual Homing Outdoors
Navigation has been defined as "the process of determining and maintaining a course or trajectory to a goal location" (M Franz et. al. (2000)). Homing is one form of navigation. In typical visual homing schemes, an agent stores a snapshot at a goal location. Later, when returning to this location, the agent compares its current view with the snapshot and uses the discrepancy between the two to generate a homing vector.
In outdoor environments, variable illumination and the movement of nearby objects will cause the view at the goal position to change dramatically over time. Until recently, few homing schemes accounted for these changing conditions. Insects are clearly able to solve the homing problem outdoors. In fact, many homing methods have been inspired by findings in insect - particularly honeybee - ethology. It should be possible for robots to do the same.
I propose to equip a mobile robot to home in outdoor environments. I will explore the image pre-processing techniques required to mitigate the effects of variable illumination and landmark movement. I will also seek to determine where snapshots should ideally be taken, and when.
back to topPaul Newman, Mobile Autonomy in Unknown Environments - the SLAM problem
This talk concerns the following long-term goal:
"To enable one or more autonomous vehicles to operate indefinitely in an a priori unknown environment with bounded uncertainty in position".
This is one of several "holy grails" of mobile robotics and is referred to as "SLAM" - Simultaneous Localisation and Mapping. It involves the fusion of several engineering domains and addressing some tough, pervasive questions. Firstly there are the estimation-theoretic issues. How can uncertainty in sensing, location and environment be managed in a time-tractable yet consistent fashion? Can it be made to work in workspaces of unlimited size? Then there are sensing and perception issues. How can the environment be sensed and how should measurements be interpreted?
Despite obvious difficulties, the applications for SLAM are numerous and easy to appreciate. The increase in mobile autonomy afforded be robust SLAM would be immense. SLAM is of particular importance in domains that, because of safety or technical reasons, make a human presence impossible. Typical examples include tasks in deep subsea, polluted / radioactive and topically, Martian settings.
This talk will offer some answers to the problems impeding a substantive SLAM deployment. It will introduce a new constant time SLAM algorithm, which is independent of workspace size. Progress will be illustrated via a series of recent land and sub-sea experiments involving the deployment of various autonomous vehicles in a priori unknown environments.
back to topTim Taylor, Review of the HYDRA project
I will present a review of the HYDRA project. HYDRA is a 3 year EU Fifth Framework-funded project concerning self-reconfigurable robotics. The project ends in December 2004. The talk will concentrate on the hardware and software developed by the Edinburgh HYDRA group, but I will also describe the achievements of the other partners in the project (Henrik Lund's group at the University of Southern Denmark, and Rolf Pfeifer's group at the University of Zurich). I will place the work in the context of research in self-reconfigurable robotics being conducted by various other groups around the world.
back to topRichard Reeve, Hearing aids for the field cricket
A collaboration with some human audio and hearing aid researchers has recently borne fruit in producing a new, improved model of the cricket auditory system. I'll be showing the latest results hot off the press since I just came back at the weekend from characterising the new system in Sydney. I will also discuss its remaining deficiencies, and how we intend to further improve it to make it into as definitive and complete a model as we can hope to achieve.
back to topDavid Stocks, Chiroptera pinnae: modelling, simulation and evolution
Chiroptera (bats) have a highly evolved sonar system which would be the envy of any forward thinking military. The CIRCE project aims to create a functional replica of such a sonar system, using a biologically rather engineering inspired approach.
My work within the consortium has focused on design of pinnae, the fleshy part of the outer ear which plays and important role in shaping the sound received by the bat. I will discuss the process of evolving pinnae with respect to the task of target localisation. This involves modelling using the implicit geometrical functions provided by VTK, and numerical simulation using our new, parallel, finite element based simulator being developed in conjunction with the Maersk Institute of Production Technology.
back to topGraeme A Jones, The Semantic Landscape of a Monitored Scene
The accuracy of object tracking methodologies and the subsequent interpretation of the monitored scene can be significantly improved by utilising knowledge about the monitored landscape. Such scene knowledge includes the homography between the camera and ground planes, methods of constrain the dimensions of appearance models, the depth map to enable reasoning about static occlusions in the scene, and the activity landscape identifying the principle routes and entry and exits zone. To facilitate plug and play functionality, this scene knowledge must be automatically learnt. Given a sufficient length of time, observations from the monitored scene itself can be used to parameterise this semantic landscape.
back to topJose Carlos Bins Filho, Automatic Control for Computer Vision
On the last 2 decades there has been a great amount of work on the development of generic image understanding systems so they can be used as shells to facilitate their application on different image understanding problems [Crevier97]. We have still to achieve that goal, but the characteristics that those systems are beginning to get clearer. Among the problems that prevented the achievement of such systems are the complexity and computational cost of low-level and intermediate level operations necessary for image understanding. This problem is smaller nowadays since there are a great number of off-the-shelf Computer Vision libraries of operators available on the Internet. These libraries provide the necessary operators for many understanding tasks, but they need a method to automatically select which libraries to execute, over which data and with which parameters. That means an intelligent controller capable of reasoning and acquiring knowledge about the task, the operators, etc. Although most researchers agree on the importance of studying and formalizing image understanding controllers, there is no agreement, though, on how that should be achieved. Some [Bulitko03] and [Draper96] advocate addressing the control problem as a separate problem from the image understanding task, while others [Crowley95] advocate that vision systems can not be designed in isolation from the task.
back to topDavid Tweed, Modelling and inferring individual and group behaviour.
In computer vision an important task is moving from conclusions relating to observed images (e.g., image features) to conclusions about the actual scene. Sometimes this process is trivial, but often this is a complex step because of inherent ambiguity and our limited understanding of the image and scene. An important example occurs in the CAVIAR project of relating moving object bounding boxes and other features from surveillance images to an interpretation of what the individuals in the scene are doing, both individually and when they interact in groups.
In this talk, I will discuss how behaviour and its relation to image features might be modelled and how inference can be performed using such models. Although motivated by vision based problems, the topics discussed may be more generally relevant to extracting "behaviours" of more general processes from low-level features.
There will also be an opportunity to observe the skilful method acting of the researchers on the CAVIAR project.
back to topTurgay Temel, Classification of Bisonar Targets
An improved low-pass filter description to be employed in biosonar signal processing with a cochlea model (for bats) is proposed and examined. It is compared to conventional model using a modified dicscrimination analysis and both are tested for assesing the performance of a given filterbank channel assuming no model yet to be developed.
In order to quantify the performance improvement, various probilistic classifiers, i.e. parametric and non-parametric models, are examined with a neuro-spike coding algorithm for echo signals from a number of different tree types. Performance improvement is exemplified with experimental results based on single-echo (single-shot) and sequential decision paradigms. Results indicate that proposed model brings in superior classification performance. Some unsupervised clustering techniques are currently being underway.
back to topMarc Toussaint, Learning representations: The genome as a self-organized representation to model phenotypic variability -- and how that relates to things like ICA
I will first briefly introduce to some of my work on the evolution of genetic representations. The key point is that when there exist multiple genetic representations of the same phenotype, then there is an implicit selection pressure between them which is related to the Kullback-Leibler divergence between their phenotypic variability and the fitness distribution---genetic representations evolve according to how well they allow to model a desired phenotypic variability. That's the way evolution learns about the structure of the "optimization problem".
I will then discuss a broader point of view by considering a general representation problem. The choice of genetic representation and similar problems can be viewed as special cases. I'll argue for incremental approaches to learning representations. When the data are strings, grammatical (hierarchical, modular) representations can be incrementally build up. But what if the data is continuous sensori-motor data?
back to topErnesto Andrade, Analysis of Anomalies in Crowd Behaviour: Local Features versus Global Knowledge
Visual surveillance in dynamic scenes is currently one of the most active research areas in computer vision. Among the applications of surveillance systems are access control in special areas, human identification at a distance, interactive surveillance with multiple cameras, crowd flux statistics and congestion analysis, and detection of anomalous behaviour.
Inside the BEHAVE project we are investigating people's behaviour for surveillance purposes. We concentrate on the detection of anomalous behaviour, which will be of interest to a security operator monitoring the scene. The analysis of such behaviours range from an individual interacting with elements of scene, to small groups and crowd dynamics.
This talk is going to discuss the upper end of this spectrum, crowd dynamics, where traditional approaches based on tracking of individuals and their body parts fail. We will discuss how we intend to solve this problem by modelling people's behaviour based on the optical flow of typical crowd scenes.
back to topAndrew Wallace, Photons, Bursts and 3D Scene Description
In this seminar, I shall try to give an overview of recent work at HWU in 3D image acquisition and analysis, from single photons and bursts of illumination at the front end of the system to object recognition and scene reconstruction at the back end.
Triangulation is still the most common form of 3D image acquisition and has led to a number of successful products that exploit either the geometry between a laser and a camera or between two or more cameras and a projected pattern to give a partial of full 3D surface model of the object. However, it has many disadvantages, notably occlusion, lack of scale flexibility, and the need for accurate geometric calibration. In particular, it has not been easily applied to acquire images of distant objects.
In the first part of the talk, I shall describe recent and continuing work to develop 3D sensors based on single photon counting, measuring reflections from targets at ranges of up to 17Km. I shall also describe briefly the work we are doing to interpret data from flash lidar systems. In each case, the requirement is to acquire detailed 3D surface models of objects quickly and from a distance.
In the latter part of the talk, I'll mention briefly past work on object recognition and how that is being developed or adapted to process the types of images acquired by these sensors. I shall also review work on semantic boundary detection and how that has developed into a current project to reconstruct 3D scenes, surface geometry and reflectance and illumination models from concurrently acquired 2D and 3D data.
back to topThor List, Thesis Proposal: Tracking Close Proximity Encounters
The research proposed in this talk will seek to investigate the problem in Computer Vision of tracking objects during encounters and interaction in a scene monitored by a single static camera. Humans do this exceptionally well because we have very high resolution stereo vision and a well calibrated low-level vision system to identify and separate objects, even when some are partially or fully occluded. We also rely heavily on strong priming from our higher-level reasoning system to explain what is going on in the scene, and we base this on a continuous analysis of current and recent events and on expectations about the scene estimated from context as well as on perception of the 3D world. We have a unique ability to re-evaluate a situation based on new evidence and re-estimate previous assumptions based on current probability calculations.
I hope to mimic the human approach by collecting detailed information from each moment and keeping this in hierarchical spatio-temporal structures, which can then be traversed, analysed and modified by a higher-level reasoning system as new information becomes available, thereby backtracking and re-evaluating previous estimates of tracked objects and their interaction. To do this I propose to construct a scale-invariant mathematical scene representation, traversable by multiple top-down processes, analysing, annotating and augmenting the existing information.
back to topJan Wessnitzer, Introducing the SPARK project: perception for action in insects
I will briefly introduce the aims of the SPARK project. SPARK is a 3-year Sixth-Framework-funded project investigating perception for action which started September 2004. The talk will briefly introduce the consortium involved in the project and present its main aim of representing and unifying biological senses under the framework of SPatio-temporal ARray Computing (SPARC) structures. IPAB's involvement in the project so far is to derive characteristics of the SPARC architecture from biological evidence and current theories of perception for action. The talk will mainly focus on perception for action in insects and aims to present some underlying ideas and general principles.
back to topScott Blunsden, Introduction to the BEHAVE project
This seminar will serve as an introduction to the behavioural understanding aspects of the BEHAVE project. Within this overall project there are two main themes. The first focuses upon the analysis of dense crowd scenes where there may be hundreds of people in view at once.
The second major area of investigation, and the main subject of this talk is in the understanding of interactions between small groups of people within the visual scene. Here the aim is to investigate the question of identifying when dangerous situations may occur. For example we concentrate our efforts upon trying to capture the behavioural aspects of (for example) a pre-fight situation which may display characteristics which distinguish it from other, normal, behaviour. The hope is that it is possible to identify dangerous situations pre-event rather than identifying them whilst they are in progress.
Work to date will be presented along with the direction that is currently being taken with the project.
back to topJay Bradley, Adapting Reinforcement Learning for Computer Games: Using Group Utility Functions
Group utility functions are an extension of the common team utility function for providing multiple agents with a common reinforcement learning signal for learning cooperative behaviour. In this paper we describe what group utility functions are and suggest using them to provide non-player computer game character behaviours. As yet, reinforcement learning techniques have very rarely been used for computer game character specification.
I'll show the results of using a group utility function to learn an equilibrium between two computer game characters and compare this against the performance of the two agents learning independently. I'll also show that we can learn equilibria between a team of two characters and a single character.
I'll also highlight some implementation issues arising from using a commercial computer game engine for multi-agent reinforcement learning experiments.
back to topToby Breckon, 3D Surface Completion using Non-parametric Techniques
As the requirement for more realistic 3D environments is pushed forward by the computer {graphics|movie|simulation|games} industry attention turns away from the creation of purely synthetic, artist derived, environments towards the use of real world captures from the 3D world in which we live.
However, common 3D acquisition techniques, such as range scanner and stereo capture, are realistically only 2.5D in nature - such that the backs and occluded portions of objects cannot be realised from a single capture.
We consider the completion of the hidden or missing portions of 3D objects after the visible portions have been acquired with 2.5D (or 3D) capture. Our approach uses a combination of global surface fitting, to derive the underlying geometric surface completion, together with an extension, from 2D to 3D, of non-parametric texture synthesis to complete localised surface texture relief and structure.
Additionally we show how this technique can be extended to increasingly available colour 2.5D/3D range data. Through this combination and adaptation of existing completion techniques we are able to achieve realistic, plausible completion of 2.5D range captures.
See example
back to topTimothy Hospedales, Signal processing in the vestibulo-ocular reflex: implications of intrinsic excitability and asynchronicity
The Vestibulo-ocular reflex is concerned with gaze stabilization in response to head/body movement. Its exhibits near linear frequency response, very low latency and high fidelity. How are these properties achieved simultaneously in a neural system? In this talk I will try to answer this question by describing our recent work relating theoretical results on population coding to the unusual properties of medial vestibular nucleus neurons invitro and invivo.
back to topMykel Kochenderfer, Integrating Planning and Learning using Adaptive Region Graphs
This talk will introduce a new system called the "Adaptive Region Graph System" (ARGS) that integrates model learning and reactive plan construction in real-time autonomous agents. This system is designed for solving problems that are more general than Markov decision processes, allowing non-finite state spaces and durative actions. ARGS consists of interacting asynchronous processes that read and modify a data structure that dynamically partitions the state space into regions using a decision graph and models transitions using a region graph. I will demonstrate ARGS in two illustrative domains.
back to topRuth Aylett, Running scared, learning to care: two affective agent architectures
Affective processing is a growing research area, as for example shown in the new EU network Humaine. However the role that affect should play in agent architectures is still an open question, where the low-level neurophysiologically-oriented community and the high-level appraisal theory community are tackling questions in rather different ways. We consider two agent architectures: one for panicking sheep and the other for empathic characters in education, and consider if the twain can meet.
back to topPaul Crook, Learning in a state of confusion
This talk is an overview of my PhD thesis. My thesis considers the application of reinforcement learning to agents who are embodied, embedded and situated in a world who's state, as far as the agents' sensors are concerned, is partially observable. That is to say the world can be modelled as a partially observable Markov decision process (POMDP). The aim of our work is to show that reinforcement learning coupled with active perception can reliably learn deterministic observation-based policies which successfully achieve a range of tasks in partially observable worlds. A deterministic observation-based policy is one where the agent forms a simple association between actions and observations such that it executes the same action every time it encounters the same observation. Such policies are also referred to in the literature as memoryless or reactive policies. The traditional approach to learning POMDP tasks is to use learning algorithms which try to uncover the underlying Markov states of the world. In contrast our approach learns policies that work around ambiguous observations, a strategy which we believe is more reliable and scalable.
back to topBruce Lamond, A New Integration of Photogrammetry and Laser-Scanning for Image-Based Rendering
Realistic and efficient reconstruction of manmade environments is of great interest in VR, gaming, archaeology, forensics etc. Two recent approaches in this area are passive photogrammetry and active laser rangefinding. Photogrammetry uses the observed perspective in a sparse set of photos to reconstruct a simple but consistent geometric representation of the scene. This can then be textured with the original images. Laser-scanning recovers a very dense discrete point cloud representation of the scene without any underlying geometry. Photogrammetry offers the ability to create photorealistic models based on a simple geometry. While being efficient, this simple geometry is inadequate where the scene becomes complex. Laser scanning offers an extremely accurate description of the scene. Unfortunately the 2.5D nature of the capture suffers from holes behind occlusions and, while futher scans can help, data sets quickly become large and alignment of scans is tricky. I will present a new hybrid approach which attempts to harness the advantages of both techniques while addressing their inherent limitations towards the end goal of photorealistic geometrically accurate scene descriptions.
back to topAnthony Ashbrook, Innovative Applications for Camera Phones using Computer Vision
The rapid uptake of camera phones has created an opportunity to deliver innovative applications that utilize computer vision to a potential audience of 215 million users (today). This new platform and audience represents a tremendous commercial opportunity as computer vision techniques more commonly associated with industrial or specialist activities are exploited in the mass consumer market. However the platform also presents a number of technical and commercial challenges that must be overcome to fully realise this potential.
In this presentation I will give a brief overview of the camera phone as a software platform for computer vision and discuss some of the innovative applications that are emerging. I will also introduce "Spellbinder", a technology for camera phones developed here at the University of Edinburgh with support from the Scottish Enterprise Proof of Concept fund.
back to topTim Lukins, Adding colour to 3D flow
The computation of a dense instantaneous velocity field that describes the displacements between two instances of a surface can be described as scene or range-flow. This form a 3D extension of the many optical-flow techniques as are traditionally based on analysing planar intensity images. A much neglected component is the use of colour, which naturally provides an additional source to disambiguate localised motion.
In this talk I will describe some derivative based approaches to calculating the range-flow, and how we can incorporate colour to further constrain the estimates. I will highlight a variety of issues in spatial-temporal filtering, colour-space conversion, and the use of synthetic data. Additionally, I will demonstrate how the techniques may be applied to 3D facial expression data.
back to topGraham McNeill, Shapes, Landmarks and Motifs
Objects are often most easily identified by their shape (as opposed to e.g. color or texture). However, it is difficult to formulate a quantitave description of shape that can be used in tasks such as medical image classification and content-based image retrieval. Here I will demenstrate that simple point matching is just as effective as the more sophisticated approaches suggested recently. In addition to this, there are techniques for speeding up point matching algorithms and the simplicity of the shape representation makes it easy to incorporate additional information. Finally, I will present some new ideas for dealing with partial occlusion and feature detection.
back to topNarayanan Unny E, Online and real time learning
Online learning is a paradigm that has been very useful in various learning scenarios involving either a large amount of data or data arriving as a stream. If the learning in such situations can be made online and real time, then it opens up a whole new avenue for faster and interactive applications.
In this talk, i will be looking at various issues involved in online learning. The need for a localised learning in an online setting and how the various algorithms like locally weighted projected regression (LWPR) and conditional mixture models try to solve these issues.
back to topGeorgios Petkos, Multiple models for motor control under different contexts
Robots are often required to operate in different environments and interact with objects that have varying physical properties. Thus, they have to be controlled by mechanisms that are able to compensate for different conditions in the external world and provide appropriate motor commands under different contexts. It has been argued that humans, in order to execute accurate movements under different contexts, are using a set of models, each of which is appropriate for a different context. Similar ideas have been proposed for artificial systems as well.
In this talk, I will discuss the basic motivation for using multiple models for motor control, present some of the work that has been done in the field, highlight some issues related with using multiple models for control and sketch some of the work that I plan to do.
back to topMark Payne, A Model System for Investigating Sensory Integration
This talk covers the background to my PhD project.
Organisms and artificial organisms may be endowed with multiple sensory modalities. How information from these modalities is combined in the pathway from sensing to action is crucial to an organism's survival and success. The number of possible schemes for this combination afforded by neurons and silicon is limitless. Learning how this combination can be done effectively by studying the products of evolution has the key advantage that nature has dealt with the complexity of the world for us. In this talk I explain why the cricket is a good creature for the study of sensory integration, what behavioural and neurophysiological data there is to go on, and how I intend to expand this into a robot model which can be used to generate and test new hypotheses about insect sensory integration.
back to topScott Blunsden, Identification of Group Interaction
The aim of this research is to use computer vision techniques to investigate the detection and classification of group interactions and behaviour. The goal is to automatically recognise if people within an image sequence are interacting, using only visual cues. Once interaction has been established the type of interaction will then be classified. A simple example of this interaction could be three people meeting or following one another.
Within this talk current methods for modelling group interaction will be presented along with their current limitations. A proposed solution is then presented, consisting of a (modified) coupled hidden Markov model, which is learned on-line to establish interactions. Classification of the specific type of interaction is performed pairwise for each of the people which have been identified as interacting.
back to topJose Vazquez, Fusion of Triangulated Sonar plus Infrared Sensing for Localization and Mapping
In this talk I will cover the research done in the past months on the area of localization and mapping for one robot, where experiments with a real robot platform where performed and validated in simulation in larger and more complex environment.
The approach incorporates information from sonar and infrared sensors mounted on a rotating platform to obtain feature-based stochastic maps of the environment. The purpose is to reliably determine the position of a robot and the features in its environment using low cost sensors.
Line and corner features are extracted from the sonar sensors by means of triangulation from multiple vantage points, while line features are extracted from the infrared sensors in separate processes. RANSAC-based approaches are used to extract the features from sonar data and from infrared data. An extended kalman filter is used to update the position of the robot and the features. The addition of infrared data to sonar data makes maps more accurate and compact.
back to topNils Roeder, Replacing the Brain Brick
The Brain Bricks from the Intelligent Autonomous Robotics (IAR) course are not getting any younger and IPAB is slowly running out of replacements. I am currently working on replacing the micro-controller (brain brick), the operating system and the current IAR programming language CPL. I will present the new Brain Brick and discuss how it will change the existing IAR course.
back to topHugo L Rosano, Biologically Inspired Six-Legged Robot
For robots working and exploring in unknown and rough terrain the use of legged locomotion is advantageous because their movement is less constrained by the shape of the surface on which they have to travel. A lot of research has been done in this area; nevertheless, up to now all of the simplest walking animals are superior to the most sophisticated artificial machines built.
We are particularly interested in the stick insect as a model because it is a slow walker accustomed to highly unpredictable environments. Neurophysiological data indicate that the walking pattern generator of the stick insect depends extensively on sensory information for patterning motoneurons. Furthermore, it is based on a decentralised architecture, in which legs coordinate with each other by means of intersegmental connectivity.
Our goal is to incorporate fundamental control and behavioural strategies of walking insects into a reliable model. Particularly we are interested in mimicking the flexibility of the single leg controller, which receives only a simplified high-level command and whose communication with other legs is minimal. For this to happen, it is necessary to have legs with complete feedback from the world, not simply position information. Furthermore, the task legs have to accomplish with each step needs to be more general and highly sensory dependent. The model will allow an artificial machine to move through complicated landscapes and adapt to various obstructions as well as allow biologists to test new hypotheses.
back to topTarik Rahman, Recovering Surface Properties for Object Manipulation in Indoor Scenes
A physically realistic scene can be created graphically by using algorithms and mathematical models to compute the
1. lighting characteristics
2. surface reflectance properties of objects
3. geometry of a scene
to form an image. When a photograph is taken of a scene, the RGB data in the
digital image can be used to work backwards and inversely calculate the above unknown characteristics of a scene provided there are some known values. For instance, if the geometry, reflectance properties and camera pose are known for an image, inverse lighting methods can be used to calculate the lighting characteristics for the scene. When this is done the scene can be relit with different lighting conditions. However it is not quite as simple as this because there are a number of assumptions to be made for different inverse rendering problems such as assuming all surfaces are diffuse in a scene, or all surfaces are isotropic etc. This talk will cover the background, problems and assumptions of inverse rendering and recovering surface properties from indoor scenes. In order to insert a synthetic object into a scene image-based lighting using an environment map is used. I will also talk about evaluating high dynamic range environment maps in the lighting simulation program, Radiance.
Rowland Sillito, Towards Anomalous Motion Trajectory Identification
My initial PhD topic was to investigate models of visual attention for video sequences - this has evolved towards an investigation of systems for identifying anomalous motion trajectories in a surveillance context.
I shall discuss the underlying reasons for the (hopefully non-anomalous) trajectory that my topic has followed, and then outline the background and specific goals of my proposed project, including: the motivation for semi-automated video surveillance; an overview of existing systems for trajectory analysis, and my proposed contribution to this area.
back to topChris Melhuish, Energy Autonomy in robots
For a robot to behave truly autonomously it will need not only to use its energy in an effective way but also extract this energy from its environment. This requires the robot to convert energy from natural raw materials and also deal with replenishing reserves and waste management. A major barrier to the widespread deployment of autonomous robots in remote areas (away from power utilities) is the availability of energy. Research at the IAS lab at the University of the West of England is looking into ways in which energy can be extracted from raw substrate in the environment - ie robots which feed. The group has constructed a mobile robot which, by employing novel microbial fuel cell (MFC) technology, can convert sugar into electricity, move to a target and transmit sensor information (Ecobot I). The group is now concentrating on the generation of sugars and equivalent input molecules from raw feedstocks such as plant material and insects. The group have constructed Ecobot II which has used 8 dead flies as the food and the robot has continued to function for over 10 days (albeit slowly!). This is the first robot in the world to run on unrefined biomass.
back to topPeter Ottery, Using Differential Adhesion to Control Self-Assembly and Self-Repair of Collections of Modular Mobile Robots
In this talk I will summarise the work I have carried out over the course of my PhD. The focus of this work has been the development of distributed control method which allows a collection of independently mobile robotic units, with two or three dimensional movement, to self-assemble into self-repairing hierarchical structures. The proposed method utilises a simple model of the cellular adhesion mechanisms observed in biological cells, allowing the robotic units to form virtually bonded aggregates which behave as predicted by Steinberg's differential adhesion hypothesis. I will present the final model and discuss some of the simulation results that have been obtained.
back to topGeorgios Petkos, Learning dynamics under varying contexts
The dynamics of the environment that a system has to interact with or even of the system itself are often changing in a rapid or discrete way. In such cases, classic adaptive control methods are inadequate since they result in large errors and instability in the period of adaptation. Furthermore, if the dynamics change back and forth, readapting everytime is a suboptimal and inefficient strategy. The use of multiple models has been proposed to cope with the problem, however relevant paradigms are not satisfactory and leave a lot of questions unanswered. The goal of this research is to come up with methods that will make control possible under such rapidly or discretely changing dynamics.
In this talk I will present the background to the problem, existing approaches and the issues with them and finally I will sketch the work that I plan to do.
back to topAroosha Laghaee, Hormone Inspired Adaptive Distributed Control for LARA
LARA is a hyper redundant modular arm with all modules being identical. This is to make production more cost effective and also allow us the flexibility of adding and removing modules when desired. This means that when it is being controlled we prefer to be able to give our instructions on the basis of each modules type vs. unique identifier at that time (e.g. the "elbow" vs. "joint3").
This required flexibility follows that a distributed control system is required which could cope with the changing configuration of the arm. I will be discussing a hormone inspired approach to control and communication for LARA and explaining how this could cope with the afor-mentioned requirements.
back to topZhicheng Liu, Reasoning about actions and change in situation theoretic framework
Reasoning about actions and change in Artificial Intelligence has been in research for a long time. However current available formalisms suffer a lot of drawbacks, such as frame problem, ramification and qualication problem etc. A lot of complex solutions have been proposed to solve these problems. But these solutions have problems of their complexity and computational difficulty. These formalisms and solutions are based on the same principles.
I take into consideration situation theory which refuses to take truth values as denotations of sentences and assumes partial situations supporting infonic knowledge. Using situation theory as a logical substrate is more amenable to a computationally tractable implementation than standard model theory of predicate calculus as it emphasizes partiality. In this talk I will introduce a situation theoretic framework for reasoning about actions and change. This framework is more expressive and flexible to overcome those problems found in this area.
back to topMatthew Szenher, Expected Difference Surfaces in Natural Environments
In the past few years, a number of researchers have found that, in a variety of natural environments, the difference between a reference intensity image and images captured nearby increases monotonically with distance from the reference location. The image difference as a function of spatial distance is typically a unimodal function whose minimum occurs at the reference location. I call this function a difference surface. Zeil et al. (2003) devised a novel visual homing algorithm which took advantage of the characteristic shape of difference surfaces. I propose that the shape of the difference surface springs from a particular quality of natural scenes. This claim leads to a mathematical description of difference surfaces. I shall use this derived function to make predictions about the efficacy of visual homing in various environments and under different conditions.
back to topDarren Smith, Toward A Simulation Of Context Generalisation
I will be talking about both theoretical and practical aspects of my PhD. On the theoretical side I will explaining more about the problem of "context generalisation," which has been described as a non-elementary learning behaviour observed in fruit flies; I will outline what kinds of experiments I am working toward to test ideas about how an insect neuropil called the 'mushroom body' may contribute to this behaviour. On the practical side I will talk about the present stage of the simulation I am building - and the problems with it - which is the development of position invariant responses to visual stimuli.
back to topBob Fisher, Overview of the CAVIAR project
The CAVIAR project is investigating several topics in video sequence understanding:
This talk will give an overview of what the project has achieved so far, and where it hopes to get to in the final few months.
back to topTimothy Hospedales, Segregation & Integration in Multi-Modal Perceptual Inference: Latent Variable Models
Multi-cue or true multi-modal perception can outperform uni-modal perception because the noise processes for each cue or modality are different. Statistically optimal methods for inference when combining multiple observations are well known in the machine learning literature. However, it is only recently that these Bayesian methods have been increasingly used to understand the statistical optimality of human multi-modal integration in experimental psychophysics results. Good understanding of and solution to multi-modal integration is only part of the multi-modal perception problem. Integration in only desirable when the observations do indeed come from the same source; an assumption made by most previous work in human and machine multi-modal perception. If, however, they do not, then the observations should be segregated instead. Optimally deciding, or discriminating, between these cases is a significant, previously un-addressed, problem which will be the core part this proposed PhD research programme.
Using Bayesian networks as a statistical formalism, we will investigate the models most appropriate for encompassing optimal integration, segregation and discrimination (ISD). Once the best models are selected and understood, we will apply them to two main applications: Constructing a robust multi-modal, multi-object tracking system for machine perception and understanding human experimental psychophysics results.
back to topNarayanan Edakunni, Localised online learning in a probabilistic framework
Online learning is an important paradigm of learning where it is required to learn from data on the fly. This paradigm of learning comes with its own unique set of problems including the inability to adapt to changing input distribution. Locally Weighted Projection Regression (LWPR) is a non-competitive localised online learner that successfully deals with this problem of negative interference, but suffers from the problem of being sensitive to parameter initialisation. This serves as the motivation to come up with a learner set in a probabilistic framework. Hence, the aim of this research will be to come up with a probabilistic model for an online localised learner.
In this talk, i will try to motivate the need for online learning, discuss the existing approaches to localised learning including the drawbacks of these models and finally propose a new probabilistic generative model called the random coefficient model that promises to serve as a good localised learner.
back to topEric McKenzie, Towards a new method for Level of Detail in Computer Graphics
This talk will briefly describe some work that has been ongoing here during the last few years. Recent analysis of the work is leading to a proposed line of research. Level of Detail research at Edinburgh goes back around 10 years. I will review that work and outline a little of the history of that time which includes the creation of EdVEC. An investigation of scale-dependent curvature flow algorithms showed that the technique could not be directly applied to level of detail. I will briefly decribe the technique and its real application. I will describe our analysis with level of detail control in mind and show how we might modify the algorithm for level of detail use.
