•  
•  
•  
•  
•  
•  

Seminar Abstracts

Hannes Saal,Spatiotemporal distribution of tactile information across the human fingertip

IThe tactile system of the human fingertips provides information which is crucial for dexterous object manipulation and it does so quickly and reliably. Decoding of relevant sensory information from the neural responses of tactile primary afferents involves identifying features like curvature of objects in contact and direction of fingertip forces.

An important question is which neural codes could transmit rich tactile information rapidly and whether they are used in the tactile system. Another question is how the spatial arrangement of the tactile receptors on the fingertip as well as the mechanical properties of the skin influence the neural response. Taken together, we ask for a spatiotemporal characterization of the tactile information.

In this talk I will discuss how information theory can be applied to tackle these questions and present some results of my analysis of neural data from the human tactile system.

back to seminars

Matt Howard,Direct Policy Learning from Constrained Motion Data

Recent advances in the study of controllers for redundant manipulators have resulted in a generic constraint-based framework both for dynamic and kinematics-based control. An important part of this framework is control in the nullspace of constraints. Commonly, this is done using some control policy either hand-designed by the robot engineer or assumed given in schemes for learning control.

n my talk I will discuss two novel methods for determining nullspace policies by estimating them from movement data. In the first, given constrained movement trajectories, we make local models of a potential function and perform alignment of these to create a globally consistent policy model. This method makes a rather strong assumption that the policy is conservative (potential based), but has the advantage of being very data-efficient. In the second we introduce a novel risk functional that allows us to make a meaningful comparison between the estimated policy and constrained observations. This allows us to model any arbitrary policy directly from constrained observations using several regression techniques trained with the modified error function.

I will present results demonstrating the methods on systems of varying complexity including kinematic data from the ASIMO humanoid robot.

back to seminars

Matt Snel,Evolution of Valence Systems in an Unstable Environment

In artificial systems employing reinforcement learning (RL), reward was traditionally implemented by having a pre-designed reward function provide the agent with a scalar reinforcement signal. It is, however, debatable whether an agent that has to rely on a pre-designed reward function to learn can truly be called adaptive and autonomous. Several approaches for moving RL away from pre-designed reward functions exist; for example, Evolutionary RL (ERL), that inspired the approach that I will present in this talk, and, more recently, Intrinsically Motivated RL (IMRL).

As in ERL, we evolve agents using an actor-critic learning scheme based on homeostasis of internal drives like hunger and thirst. In particular, our paper compares the performance of drive- versus perception-based motivational systems in an unstable environment. We investigated the hypothesis that valence systems (systems that evaluate positive and negative nature of events) that are based on internal physiology will have an advantage over systems that are based purely on external sensory input.

Results show that inclusion of internal drive levels in valence system input significantly improves performance. Furthermore, a valence system based purely on internal drives outperforms a system that is additionally based on perceptual input. I provide arguments for why this is so and relate our architecture to brain areas involved in animal learning.

back to seminars

Jan Steffen,Dextrous Grasping and Manipulation Using Manifolds

In dextrous hand control, the implementation of manipulation movements still is a complex and intricate undertaking. Often, a lot of object physics and modelling effort has to be incorporated into a controller working only for a restricted task specification and performing quite artificially looking movements.

In this talk, starting from a representation for dextrous grasping - the Grasp Manifold - I motivate and present adaptations which enable a modified representation - the Manipulation Manifold - to robustly represent manipulation movements. We use manifolds of hand postures embedded in the finger joint angle space which are constructed such that manipulation parameters including the advance in time are represented by distinct manifold dimensions. This allows for simple purposive navigation within such manifolds. I present the first steps towards the construction of such manifolds using the Unsupervised Kernel Regression (UKR) and the way of applying it for manipulation in the example of turning a bottle cap in a physics-based simulation.

Finally, I will give a short overview of the ideas and problems we would like to address during my research stay in Edinburgh in order to realise a more unsupervised construction of the presented movement manifolds.

back to seminars

Djordje Mitrovic,Adaptive Optimal Feedback Control for high-dimensional Movement Systems

In recent years Optimal Feedback Control (OFC) has become the predominant motion generation strategy for biological movement systems. OFC not only explains most motion patterns observed in human reaching but also allows a mathematically coherent formulation. This makes it an appealing theory for modeling volitional motion of biological and artificial systems.

Most OFC models make simplifying assumptions (linear dynamics model and quadratic cost function) due to the computational limitations OFC methods impose. In this proposal I investigate OFC for highly nonlinear and redundant systems, based on iterative optimal control methods. So far these methods relied on an analytic form of the system dynamics, which may often be unknown, difficult to estimate for more realistic control systems or may be subject to frequent systematic changes. I present a novel combination of learning a forward dynamics model within the OFC framework. Utilising such adaptive internal models can compensate for complex dynamic perturbations of the controlled system in an online fashion. This allows us, for the first time, to study OFC from an adaptation perspective and its link to biological motor control.

In my talk I will introduce this adaptive OFC framework and present results from several adaptation experiments in simulation. I further will motivate my planned future research towards a hardware implementation and a biological interpretation of my adaptive OFC model.

back to seminars

Giorgos Petkos,Learning dynamics for robot control under varying contexts

High performance control algorithms require an accurate model of the system's dynamics. Learning dynamics is an attractive alternative when accurate analytical derivation is not possible. However, dynamics may exhibit nonstationarity due to interaction with different environments or objects. We refer to the unobserved factors that affect the dynamics as the context of the dynamics. Under nonstationary dynamics, the dynamics model needs to be adapted whenever the context of the dynamics changes, otherwise there will be large tracking errors. In this talk, we examine ways to improve performance by reusing knowledge obtained by experiencing the dynamics of a set of contexts.

We consider two classes of scenarios. In the first the dynamics may switch between a finite set of contexts. In that case we use a set of learned models for each of the contexts and switch between them accordingly. We formulate this as a probabilistic discrete latent variable model. In a more complicated scenario with continuous, possibly infinite number of potential contexts, the use of a set of models may not be viable and generalization to novel contexts is required. To tackle this problem, we reformulate the probabilistic discrete latent variable model as a continuous latent variable model.

back to seminars

Mark Payne,Mushroom Bodies and Motor Control

The mushroom bodies of the insect brain are highly prominent, highly structured regions which have been researched for many years. In common with parts of the mammalian brain which share these characteristics (e.g. the hippocampus, the cortex and the cerebellar cortex), the functional role of the mushroom bodies is not clearly understood. Compelling evidence has shown that they are involved in olfactory conditioning, but ablation studies and genetic mutants also point to a role in suppression and termination of locomotory bouts.

My own work into sensory integration in the cricket has focussed on the role of efference copies and sensory predictions, in particular the suppression of turning due to self-generated optical flow. In this talk I will discuss the components that are necessary in a system that performs such a task, how these might map onto the mushroom bodies, and present my attempts to implement the system in a spiking neural network for controlling a robot with sound-tracking and optomotor behaviours.

back to seminars

Michael Mangan,Navigation in insects using multiple place memories

In this talk I shall provide a brief overview of my current behavioural study in which I seek the mechanism underpinning idiosyncratic route formation in the desert ant Cataglyphis ibericus. I shall also present results from a previous behavioural study in which the ability of the cricket Gryllus bimaculatus to return to a hidden target location using only visual cues was tested.

I shall then compare the performance of different classes of homing algorithm in the various experimental paradigms outlined above. Furthermore, I shall discuss the limitations of such models and offer a novel solution using a classical neural network architecture. Preliminary results using this technique shall be presented and future work with the aim of forming a general model for route navigation in insects shall be discussed.

back to seminars

Flavio Prieto (National University of Columbia),Inspection of 3D Deformable Parts Using Radial Basis Functions

In industry, one of the most common schemes to perform automated inspection tasks consists of matching a design represented by a 3D CAD model with 3D data measured by geometric sensors such as laser scanner or industrial CT. To carry out this comparison it is necessary to first align or register both the CAD model with the part 3D measurements. Usually, the part to be inspected can be considered to be rigid where a simple rigid body transformation is enough to achieve alignment with the CAD model. However, modern manufactured parts are becoming more and more flexible due to new materials such as composites. Traditional inspection system requires in this case that the part have to be fixed in place using clamps. This process is time-consuming and difficult to automate. What is necessary is an inspection system capable of applying to the CAD model a more general transformation that includes the deformation of the parts to the traditional rigid body assumption. In this work, we explore the use of Radial Basis Functions (RBF): Gaussian, multiquadrics and inverse multiquadrics, as a method to represent the deformations, required for the models registration during the inspection processes. In order to do this, a comparison between the deformation obtained with the Radial Basis Functions and the one obtained using Finite Elements analysis is presented allowing us to calibrate the accuracy of the inspection process.

back to seminars

Tom Larkworthy,Self-Reconfiguration Planning Heuristics

Planning algorithms for metamorphic systems based on search require heuristics because of the massive state space. The most popular heuristic has been the optimal assignment heuristic. This is an application of the well known combinatorial problem of optimal assignment. It can be calulated in O(n3).

I will present a suboptimal dynamic persistent version of the heuristic which can be calculated incrementally a cost of around O(n log(n)) (I have not worked it out yet, but its defiantly below n2). Although it arrives at sub optimal solutions to the optimal assignment problem, which translates during search to more states being evaluated during planning, the reduced computational cost more than makes up for this.

I will also present a very different form of heuristic which compresses a reconfiguration state into a fixed size vector representation. This can be calculated incrementally in linear time. The new heuristic performs worse than the suboptimal assignment heuristic when used in a greedy search. However, the fixed size vector representation allows more sophisticated meta-heuristic searches to be applied effeciently. Early results show that in some situations rapidly exploring random tree based planning using this new heuristic can outperform greedy search.

back to seminars

Jafreezal Jaafar,Reactive Behaviour In Game Design Using Fuzzy Logic

Computer games are programs that enable a player to interact with a virtual game environment for entertainment and fun. Each game has its own strategy, action, curiosity, action, challenge and fantasy that make each game unique and interesting, which can essentially motivate games players. In this presentation, the classic game Pacman was chosen to demonstrate a behaviour based system using fuzzy logic. In earlier implementation of the game, the ghost logic did not realistically adjust to user skill and movement. For instance, the ghosts did not move toward the areas where the Pacman needed to complete the level. While this could be done with classic logic, fuzzy logic can provides a better way for a system to deal with the often ambiguous data required to implement such behaviours. In addition, this type of system allows rules to be easily added to increase the opponents? intelligence further. For these reasons, fuzzy logic has been chosen as the basis for the intelligent control of the ghosts? behaviour. I will present the design and implementation of a real-time fuzzy-based system for an interactive game. The chosen game is a remake of Pac-Man in which the opponents are BDI-style intelligent agents. The components of the system and the methods used in fuzzifying the game?s rules and variables are also discussed. Finally, will discuss the results observed in the implementation of the game along with a comparison to classical design methodologies.

back to seminars

Theodoros Damoulas (University of Glasgow),From Automated Currency Validation to Protein Fold Recognition: Probabilistic Multi-Class Multi-kernel Learning

In diverse machine learning problems ranging from automated currency validation (ACV) to protein fold prediction, we encounter the situation where multiple object descriptors are available for a possibly multinomial classification task. Specifically, ACV considers the challenging and unresolved problem of counterfeit note detection while depositing currency in an ATM that is equipped with a plurality of sensors. In an analogous manner, when predicting the structural fold of a protein multiple feature sets are available, ranging from global characteristics like the amino-acid composition and predicted secondary structure, to attributes derived from local sequence alignment such as the Smith-Waterman scores.

These problems raise the need for a classification method that is able to assess the contribution of these potentially heterogeneous object descriptors while utilizing such information to improve predictive performance.

In this talk I will present a hierarchical Bayesian multi-class multi-kernel pattern recognition machine that informatively combines the available feature groups and, as is demonstrated, is able to provide the state-of-the-art in performance accuracy on the problems considered. The full Markov chain Monte Carlo solution of the model is offered via a Metropolis-Hastings within Gibbs sampling procedure and also an efficient variational Bayes approximation is proposed.

back to seminars

Joachim Hass (University of Gottingen),Time Perception and Motion Control - Two Sides of the Same Coin?

Time is a very important dimension of our life, both for perceiving stimuli changing in time, such as speech or music, and for precisely timed coordination of movements. Currently, it is debated whether these two domains rely on common or distinct neural mechanisms. In this work, we investigate the influence of a motor task on a simultaneously performed time perception task. Our results show that this influence is not limited to a simple impairment of performance, but is specific to the state of motion the participants are in.

Participants were required to follow an elliptic trajectory that was drawn on a screen, using the end effector of a robot arm. At specific segments of this guided arm motion, namely on the apecies of the ellipse, two short tones were presented which the participants had to discriminate according to their duration. Both the discrimination performance and the perceived duration of the tones were tested. We found that the duration of the tones where systematically underestimated at the apecies where the motion was slower and the ellipse was more curved, and also overestimated in the other apecies. On the other hand, speed and curvature did not affect the discrimination performance.

These results allows for the conclusion that time perception and motor control share some neural circuitry. This correspondence can be explained in the framework of the synfire chain model, under the assumption that at least some of the chains encode both time and motion. Changing the transmission delay of these chains e.g. by altered background activity would lead to the joint change in motion speed and PSE, with little influence on the discrimination performance.

Acknowledgement: This study was supported by a grant from the Bundesministerium fuer Bildung und Forschung (BMBF) in the framework of the Bernstein Center for Computational Neuroscience Goettingen, grant number 01GQ0432.

back to seminars

Jochen Steil,Online Reservoir Learning of movements for PA10 and ASIMO

We present an neural network approach for simultaneous learning of task and joint angle representations for target movements of redundant robots in a single coherent framework. For training we use an efficient online scheme for recurrent reservoir networks consisting of backpropagation-decorrelation (BPDC) output adaptation and an intrinsic plasticity (IP) reservoir optimization. We demonstrate that the network can acquire highly accurate inverse models and task predictions for the redundant 7-DOF robot arm PA-10 and the humanoid robot ASIMO, with excellent generalization to untrained target motions. The potential of the approach for imitation is shown by reproducing real data recorded from a human demonstrator. The talk will also include a short overview on recent activities in the new Bielefeld Research Institute for Cognition and Robotics (CoR-Lab) and the Excellence Cluster in Cognitive Interaction Technolgie -- (CITEC) hosting this research.

Edmond Shu Lim Ho,Synthesizing human interactions with topological constraints

In computer graphics, many researchers have been working on synthesizing realistic human animations. With the use of MOCAP systems, creating animation is no longer a time-consuming task. However, most of the existing work focus on scenes without close interactions of multiple avatars. In addition, it is difficult to capture such motions due to the limitations of the MOCAP devices.

Using individually captured motions to generate multi-character animations with close interactions is a practical approach. But it is expected that there will be a lot of collisions among the body segments of the characters. Existing methods can handle these problems by using collision detection algorithms and editing the positions of the colliding segments by inverse kinematics. However, these methods do not take into account the topological relationships between the body segments. Therefore, there is no guarantee that the topological relationships will be kept the same after the editing.

Previously we proposed a method to keep the topological constraints when editing individually captured motions. However, we do not allow the topological relationship to change. In this talk, I will present a new method to simulate the interactions between characters while satisfying the topological constraints. The new method can synthesize the transition motions between different topological states. This method is not limited to synthesizing multiple-character animations but also the interactions between character and objects with topological constraints.

back to seminars

Mark Harrison,Automatically Generating Options in Reinforcement Learning

The options framework is a popular way of trying to improve the performance of reinforcement learning agents.�  Options are higher-level temporally-extended macro-actions defined using a policy over lower-level actions, usually intended to achieve some specific sub-goal within the task.�  Using options allows us to pre-supplying an agent with a selection of task-specific skills, hopefully improving it's performance within a problem.�  However, it can be difficult and time consuming to manually construct options for a task, so a natural extension is to try and have the agent generate it's own options as it explores the environment.�  I will present some of the the main current methods of generating options on-line, showing that using such approaches can improve the performance of agents in simple test problems.�  I will also talk about some of the weaknesses of these approaches, and possible avenues for improvement.

back to seminars

Ioannis Havoutis,Learning global control strategies for dynamically dexterous robots

One of the major outstanding issues in robotics is the design of planning and control strategies that enable robots to be both flexible and robust. In other words, we want robots that are capable of performing numerous variations on highly dexterous tasks, under a variety of different environmental conditions. This brings up the need for global control strategies - and the focus shifts towards global realizability and lazy optimization in a somewhat adversarial environment. I will present some previous work on global control that inspires my research and identify some open questions in this area. Primarily, there is a need for efficient algorithmic techniques that can acquire three things from active exploration and limited skilled demonstration - local behaviors, their regions of applicability and a specification for composition of behaviors. This way we can build on simple ideas of local -skill level- behaviors and global -task level- control, suitable for high-level planning. I will illustrate my approach with a simple weakly actuated pendulum balancing task and then outline my current and future research efforts with the KHR-2HV humanoid in simulation.

back to seminars

Adrian Haith,A Bayesian Model of Multimodal Motor Adaptation

Sensorimotor adaptation is a process which is able to utilize observations from multiple modalities to improve performance. Previous computational models of motor adaptation have tended to neglect the role of proprioception, while sensory integration and recalibration have primarily been considered in the context of passive perception and not in the context of active goal-directed movements. I will present a unified model of multimodal motor adaptation to visuomotor disturbances in which learning is driven by optimal Bayesian inference of miscalibrations in visual and proprioceptive modalities. This model is able to accurately account for the time-course of visuomotor adaptation, as well as the shifts in visual and proprioceptive perception which occur as after effects.

back to seminars

Theo Gonos,Using Homeostatic Neurons for Sensor Self-calibration

Sensor array calibration is a major problem in engineering, to which a biological approach may provide alternative solutions. I will present a homeostatic algorithm for self-calibration of individual sensory neurons. The algorithm adjusts each sensor gain and offset by trying to maintain a constant spiking activity. This mechanism was tested on the proximity sensors of a robot performing obstacle avoidance behaviour. I will show how the robot adapts to different environments using the slow time scale effects of the homeostatic calibration, e.g. adjusting sensitivity to get through gaps in cluttered environments, changing responsiveness for sensors that do not contribute to the task, and producing new responses that prevent the robot from getting stuck in one behavioural cycle.

back to seminars

Bill Lewinger,Borrowing from Nature: Insect-like Robotics, Behaviour and Functionality

Insects are amazingly agile creatures capable of navigating difficult terrain with ease. By borrowing from nature, we can design biologically-inspired robots that benefit from insect evolution to accomplish similar tasks. This talk will feature past and present research in legged robotics regarding the Biologically-Inspired Legged Locomotion Ant robot (BILL-Ant), and additional control autonomy for the biologically abstracted Whegs(tm) robotic platform.

back to seminars

Hubert Shum,Synthesizing Dense Interactions for Multiple Avatars

In this seminar, we propose a new method to generate a realistic scene of avatars densely interacting in a competitive or cooperative environment. Using our method, the interactions among hundreds of avatars can be created in real-time. The motions of the avatars are considered to be captured individually, which will increase the easiness of obtaining the data, while the interactions of multiple characters are synthesized using artificial intelligence and machine learning algorithms.

We first present an approach called temporal expansion, which is used to predict the future state of interaction by expanding the game tree of two avatars. Then, using this method, we efficiently sample the high dimensional state space of interaction by exploring the subspace of meaningful interactions and favouring samples that have high connectivity with the others. Using the collected samples, we create the Interaction Graph, which is a finite state machine to simulate controllable continuous interaction. We also create Interaction Patches, which can be spatio-temporally concatenated to generated animation of an interacting crowd up to hundreds of avatars.

The proposed method is superior to previous motion synthesis techniques due to its power to simulate realistic interactions. On the other hand, the ability to plan the movements of an interacting crowd creates a much more realistic crowd of avatars when comparing to tradition crowd simulation approaches.

back to seminars

Sebastian Bitzer,Synthesising Novel Movements through Latent Space Modulation of Scalable Control Policies

We propose a novel methodology for learning and synthesising whole classes of high dimensional movements from a limited set of demonstrated examples that satisfy some underlying 'latent' low dimensional task constraints. We employ non-linear dimensionality reduction to extract a canonical latent space that captures some of the essential topology of the unobserved task space. In this latent space, we identify suitable parametrisation of movements with control policies such that they are easily modulated to generate novel movements from the same class and are robust to perturbations. We evaluate our method on controlled simulation experiments with simple robots (reaching and periodic movement tasks) as well as on a data set of very high-dimensional human (punching) movements. We verify that we can generate a continuum of new movements from the demonstrated class from only a few examples in both robotic and human data.

back to seminars

Paolo Favaro,Boosting Invariance and Efficiency in Supervised Learning

In this presentation I will introduce a novel boosting algorithm for supervised learning that incorporates invariance to data transformations and has high generalization capabilities. While one can incorporate invariance by adding virtual samples to the dataset (e.g., via jittering), we adopt a more efficient strategy and work along the lines of vicinal risk minimization and tangent distance methods. As in vicinal risk minimization, we incorporate invariance by introducing anisotropic smoothing along the directions of invariance. Furthermore, as in the tangent distance method, we provide a simple local approximation to such directions, so as to obtain an efficient computational scheme. Finally, we show that it is possible to automatically design optimal weak classifiers by using gradient descent. To increase efficiency at run-time, such optimal classifiers are projected onto a Haar wavelet basis. This procedure results in designing strong classifiers that are more computationally efficient than the case of exhaustive search. For illustration and validation purposes, we demonstrate the proposed algorithm on both synthetic and real data sets that are publicly available.

back to seminars

Christoph Kolodziejski,Closed loop control and behavioral learning using a differential Hebbian framework

In classical conditioning, an agent correlates an initially neutral with a behavioral relevant stimulus. This has been modeled by differential Hebbian learning. Most often that was done in an "open-loop" fashion, hence where behavior has no effect on learning. However, for an autonomous agent, its behavior will have an influence on the stimuli which it encounters and, as a consequence, also on learning. This can be emulated by closed-loop ISO-learning, an augmentation of differential Hebbian learning, where the output of the Hebbian neuron controls the agent and the loop is been closed through the environment. Here we will present an overview of the ISO-learning rule and its siblings in the context of neurophysiology as well as control theory. We will provide theoretical arguments that learning and behavior are stable, which will then also be demonstrated in several robotics applications presented in the second part of this presentation.

back to seminars

Poramate Manoonpong,Neural control for locomotion of walking machines

The basic locomotion and rhythm of stepping in walking animals like cockroaches mostly relies on a central pattern generator (CPG), while their peripheral sensors are used to control walking behaviors. By contrast, in stick insects, sensory feedback serving as reflexive mechanism plays a critical role in shaping the motor pattern for adaptivity and robustness of walking gaits. Inspired by the principles of biological locomotion control, two different types of neural mechanism for locomotion control of walking machines are presented. One is called modular reactive neural control and the other is adaptive reflex neural control. Modular reactive neural control based on a modular structure design applies a CPG for basic rhythmic leg movements and motor coordination of the six-legged walking machine AMOS-WD06 while peripheral sensors of it serve to stimulate a variety of reactive behaviors, like self-protection reflex, obstacle avoidance, phototropism, sound tropism, and wind-evoked escape behaviors. On the other hand, adaptive reflex neural control is based on reflex mechanisms. It uses proprioceptors, e.g., foot contact and angle sensors, to drive dynamic walking patterns, to regulate walking speed of the biped robot "RunBot", and also to synchronize its components. In addition, learning mechanisms have been integrated, the result of which enables RunBot to perform fast dynamic walking and autonomously learn to adapt its locomotion to different terrains.

back to seminars

Roderich Gross,What properties are required for individuals that work in groups and teams

Swarm robotics draws inspiration from decentralized, self-organizing biological systems in general and from the collective behaviour of social insects in particular. In a social insect colony, many tasks are performed by higher-order entities (groups, teams, and the colony itself) whose task solving capacities transcend those of the individual participants. In this talk I present two studies that investigate what properties are required at the individual level to generate such higher-order entities.

In the first study, I simulate a system of simple, insect-like robots that can move autonomously and grasp objects as well as each other. I use artificial evolution to produce solitary transport and group transport behaviours. I show that robots, even though not aware of each other, can be effective in group transport. Group transport can even be performed by robots that behave as in solitary transport. Still, robots engaged in group transport can benefit from behaving differently from robots engaged in solitary transport.

In the second study, I report on an experimental study in which up to 12 physical robots perform a foraging task. The task requires the robots to engage in a range of different activities, including exploration, path formation, recruitment, self-assembly and group transport. Once the robots start interacting with each other and with their environment, they self-organize into teams in which distinct roles are performed concurrently. The system displays a dynamical hierarchy of teamwork, the cooperating elements of which comprise higher-order entities. The study shows that teamwork requires neither individual recognition (the robots are inter-changeable) nor inter-individual differences (the robots are identical in terms of "morphology" and "brain").

back to seminars

Stefan Klanke,Learning manifolds with unsupervised kernel regression

'During my PhD, I worked on a method for non-linear dimensionality reduction, which we called "Unsupervised Kernel Regression" (UKR) due to its origin, the classic Nadaraya-Watson estimator. In this talk, I'll give an overview of UKR and its properties, and I will describe some possible extensions to the algorithm.

In a nutshell, UKR requires very little parameters to be chosen a priori: In its simplest form, a UKR model is fully specified by the dimensionality of latent space and the choice of a density kernel, and it can be regularised automatically by using leave-one-out cross-validation (without additional computational cost). The model is then fitted by gradient-based optimization, that is, the low dimensional coordinates (latent variables) and a mapping from latent space to data space fall out by minimizing some error criterion.

UKR can be extended 1) to a more general CV-scheme (aimed at avoiding unsmooth manifolds that sometimes result from LOO-CV) 2) to include loss functions beyond the usual squared error (e.g., to enhance robustness, or to tune the method to specific noise levels) 3) to a "landmark" variant which helps to reduce the computational cost 4) to Unsupervised Local Polynomial Regression, where the Nadaraya-Watson estimator is exchanged by local linear or local quadratic regression, the latter showing less bias in the presence of curvature.

back to seminars

Taku Komura,Simulating Interactions of Multiple Avatars

The main focus of our research work is to create scenes of multiple avatars interacting. One of our approaches to achieve this is to use topological relationships. Previous research about motion synthesis mainly focused on the kinematics rather than the topology.

In this talk, I will explain how topological relationships can index the postures of a single person as well as the status of two human bodies interacting, and then show some examples of segmentation / classification of human motion data based on our method. Next, I will talk about our approach to simulate fighting scenes in real-time.

Reinforcement learning is an approach to achieve real-time optimal control. However, the huge state space of human interactions makes it difficult to apply existing learning methods to control avatars when they have dense interactions with other characters.

In this research, we propose a new methodology to efficiently plan the movements of an avatar interacting with another one. We make use of the two principle facts which is common in many interactive activities of two avatars: (1) the subspace of meaningful interactions is much smaller comparing to the whole state space of two avatars, and (2) flexible postures which can switch to various motions are favored for human interactions. We use an off-policy approach and efficiently explore the search space by expanding the game tree taking into account these principles and compose a finite state machine (FSM). Then, at run-time, we compute the optimal action of each avatar by min-max search or dynamic programming on the FSM. The methodology is applicable to control NPCs in action games of fighting, ball sports or even pedestrians.

back to seminars

Subramanian Ramamoorthy,Task encoding and machine learning for dynamically dextrous robot behaviours

Robust autonomy and dynamical dexterity are two of the most important qualities we expect of modern robots (that we hope will one day go out and perform a variety of activities ranging from rescue and exploration to playing football). The combination of these two qualities has been hard to achieve. This talk will address this broad issue.

I will begin with a discussion of what it means to be robust, autonomous and dynamically dexterous; why we care and why many of the existing methodologies do not achieve the desired objectives.

Then, I will outline a specific approach to encoding these tasks and solving the planning and control problem - arguing in favour of a layered representation based on symbolic abstractions of the underlying continuous dynamics. In order to demonstrate the benefits of this approach, I will describe some design examples from my prior work.

Lastly, I will step back and discuss some major open questions that arise within this program - including conceptual issues in learning highly structured strategies from data and practical experimental challenges of achieving complex behaviours in complex robots.

back to seminars

Barbara Webb,SPARK project summary

We have just concluded the EC project SPARK (Spatial-temporal Patterns for Action-oriented perception in Roving robots) and in this talk I will present a summary of the work. I will focus most on the insect brain architecture developed at Edinburgh but also discuss some of the hardware (e.g. visual plane processors) and algorithms (e.g. using recurrent neural networks) developed by our partners.

back to seminars

Shu Lim Ho,Synthesizing human interactions with topological constraints

Creating animation of humans interacting with each other is a challenging research topic in computer graphics. With the use of motion capture (MOCAP) device, human motions can be recorded for animating virtual characters. However, it can be difficult to capture human motions with dense interactions such as those in wrestling. It is because markers will be occluded if we use optical MOCAP systems and sensors might be damaged if we use magnetic and mechanical MOCAP systems. It is possible for experienced animators to create such scenes by key-framing techniques; however, it is a time-consuming and labour-intensive task.

My research is designed to generate realistic computer animations with close human interactions by taking the topological relationships of the virtual characters into account. This research can be roughly divided into two modules: the Topology Analysis and the Motion Synthesis. The topology analysis module will extract and encode the tangle information from the captured motion data. The motion synthesis module will make use of the tangling information as topological constraints in the motion editing / synthesis process. As a result, it will be possible to create / edit close-contact motions with minimum effort by the animators. The work can be used for any movement that requires the body to be tangled with others, such as wrestling, a helper holding a shoulder of an elderly to walk; or a soldier piggy-backing another injured one. The application of this research is not limited to computer animation but also to robotics.

back to seminars

Jochen Ehnes,Projected Reality -- Augmenting the Environment with a Network of Controllable Video Projectors

In this talk Jochen Ehnes will present the concept and the ideas behind the system he developed during his PhD research at the University of Tokyo. He will start with a short introduction of related fields such as Augmented Reality (AR) and interaction on projected interfaces on tabletops as well as other room surfaces.

The aim of his work is to provide a similar degree of freedom in the sense of which surfaces (fixed as well as moveable) can be augmented as head mounted devices do. However, without users having to wear one, as they are often considered a hindrance.

By combining a video camera and projector on controllable gimbals, the projection unit can recognise certain objects, follow them and project information onto them. By using several of these projection units, geometrical limitations (e.g. the surface has to face the projector so it can be augmented) can be reduced and the overall range can be extended. Interaction with such a projection system naturally is very different from a conventional GUI. For that reason, Jochen developed the concept of 'Projected Applications' and devised an API based on that to ease the development of such applications. An important functionality of the developed architecture is to enable the projected applications to roam between different projection units, in order to always project the information from the best-suited unit. The latest version of the API also facilitates the usage of several projection units for tracking and projection onto different surfaces of an object by one application. Here the API takes care of chores like networking.

back to seminars

Sebastian Bitzer,Task Relevant Motion Generalisation

In the animation community a recent nonlinear dimensionality reduction technique (the GPLVM) has been successfully applied to represent very high dimensional, complex human motion in a very compact (2 or 3 dimensional) space. We transfer this approach to robotics where we aim to establish a link between the low dimensional space resulting from dimensionality reduction (the latent space) and the space in which we describe how well, or in which form, a given task is performed (the task space). Once we have established this relation we can generalise the exemplary movements which have been used to find the latent space to novel movements which perform the task in a different way without loosing their similarity to the original movements. Thereby we ensure the similarity of the movements by using dynamic control policies as a common representation in latent space.

Here we present first, simple simulation experiments with two kinds of robot arms which show the feasibility of this approach. The movements that we consider are straight line reaching movements which are in parallel or spread out like a star and circles and figure-8s with different diameters. We are able to reliably reconstruct test movements by interpolation in the latent space, although reconstructions for the periodic movements are not as good as those for the reaching movements. Finally we suggest improvements based on these results.

back to seminars

Prof Bob Fisher,An empirical model for saturation in classifier spaces

When assessing reported classification results based on selection of members from a database (e.g. a face database), one would like to know what is an achievable classification rate, given the noise level, dimensionality of the feature set and number of classes in the database. As best we can tell, no general results exist for this question, although many classification rates appear in different papers.

This talk presents an empirical formula for MAP classification that links the number of discriminable classes to the error rate, dimensionality of the feature data and the feature noise level.

I will also briefly present some recent highlights from the vision group.

back to seminars

Gordon Wyeth (University of Queensland, Australia),Using Models of Rodent Hippocampus for Robot Navigation

The brain circuitry involved in encoding space in rodents has been extensively tested over the past thirty years, with an ever increasing body of knowledge about the components and wiring involved in navigation tasks. The learning and recall of spatial features is known to take place in and around the hippocampus of the rodent, where there is clear evidence of cells that encode the rodent's position and heading. Many components of hippocampus have been modelled by computer simulation, and there exist some well understood computational models that exhibit similar characteristics to the recordings from the hippocampal complex.

This talk addresses two questions: 1. Can models of rodent hippocampus match the state of the art in robot mapping? 2. Can models of rodent hippocampus embodied in a robot inform biology?

The questions are addressed in the context of a system called RatSLAM which is based on current models of the rodent hippocampus. RatSLAM is demonstrated performing real time, real world, simultaneous localisation and mapping from monocular vision, showing its effectiveness as a robot mapping and localisation tool. Furthermore, some of the modifications necessary to make the models of hippocampus work effectively in large and ambiguous environments potentially raise some new questions for further biological study.

Biography:

Gordon Wyeth heads the Robotics Laboratory in the School of Information Technology and Electrical Engineering at the University of Queensland, Australia. He holds a PhD and a Bachelor of Engineering degree (with honours) in Computer Systems Engineering. He was President of the Australian Robotics and Automation Association from 2004 to 2006, and Director of the Mechatronics Engineering Program at the University of Queensland over the same period. He has served in various positions in RoboCup International Federation, and initiated the RoboCup Junior program in over 200 Queensland schools.

The University of Queensland Robotics Lab receives funding through the Australian Research Council and other government bodies, as well as industry partner SAP. The Lab has designed and constructed more than twenty types of robots, including flying robots, wall-climbing robots, high performance wheeled robots, legged robots, manipulators and a humanoid robot. The Lab's robot soccer team, the RoboRoos, have been runners-up three times in the RoboCup World Cup of robot soccer. The Lab's principal research aim is to build practical and useful robots that exploit, explain and expand models of living systems.

back to seminars

Michael Greenspan (School of Computing, Queen's University, Kingston, Canada),Is Computational Intelligence Needed to Play Robotic Pool?

Since the seminal work on computational chess of Claude Shannon and Alan Turing, computer gaming has had a long and respected tradition, and the success in 1997 of IBM's Deep Blue against world chess champion Gary Kasparov is regarded as a turning point in the evolution of computational intelligence. To date, however, no robotic system has successfully competed against a proficient human opponent. This talk will describe the development of Deep Green, a Robotic Intelligent System being developed to compete against humans at the game of pool, which currently plays 8 Ball at a better-than-amateur level. The Deep Green system comprises: a ceiling mounted gantry robot; a cue end-effector; a machine vision system; control and physical modeling algorithms; a search-tree based strategy engine; and a full-sized pool table. The background research and major design decisions will be reviewed, and recent results in automatic play will be presented. The role that strategy plays in robotic pool will also be discussed. The talk will conclude with a description of the future research problems that will be addressed to advance the system to perform at a competitive level.

Biography:

Michael Greenspan is an Associate Professor with the Department of Electrical and Computer Engineering and the School of Computing at Queen's University, Kingston, Canada. His Michael's research interests include pose determination, object recognition, and tracking especially using range data, and robotic gaming systems. Michael was the recipient of the Canadian Image Processing and Pattern Recognition 2003 Young Investigators Award, and the Premier's Research Excellence Award.

back to seminars

Michael Mangan,Desert ant navigation: are idiosyncratic routes formed through the linking of local vectors to snapshots?

Desert ants can navigate by both path integration and visually guided strategies. Recent studies have revealed that within cluttered landscapes individual desert ants travel to a familiar feeding site by an idiosyncratic route, which is statistically indistinguishable across repeated journeys. Linking visual memories, known as snapshots, to local vectors has been proposed as a mechanism by which ants may learn these long distance paths. However, conclusive evidence confirming that this mechanism underlies the desert ant trajectories has not been forthcoming. This talk shall outline how I plan to verify the snapshot to local vector model through behavioural & robot modeling studies.

NB - As I have recently given an IPAB seminar on this subject this talk shall focus more on the modeling aspects of my PhD which was only briskly covered previously. Additionally I shall incorporate my recent work on cricket spatial memory.

back to seminars

Emre Ugur (Department of Computer Engineering, Middle East Technical University),The learning and use of traversability affordance on a mobile robot

The concept of affordances was introduced by J.J.Gibson to explain how action possibilities can be directly perceived by an organism interacting with its environment. We are interested in how this concept can be used in robot control and learning. To clarify the concept and use it in robot control, we proposed a new formalism where affordances are defined as relations within the robot-environment system, and can be seen from three different perspectives; namely agent, observer and environmental. Affordance instances are represented as nested triples of the form (effect, (entity, behavior)), where the robot can learn affordance relations through interactions with the environment by forming different equivalence classes. We present three studies implementing certain aspects of the formalism on a mobile robot moving in an environment filled with different types of objects. Specifically, we show that, (1) the formation of the entity equivalence classes corresponds to the perceptual learning of affordances, (2) the formation of effect equivalence classes, followed by the formation of entity equivalence classes can lead to the development of goal-directed behaviors from a set of primitive ones, and (3) the formed equivalence classes and relations provide support for planning and deliberation.

back to seminars

Rowland Sillito,Incremental One-Class Learning with Bounded Computational Complexity

The problem of one-class learning presents itself in cases where one wishes to distinguish between members of a class for which examples are abundantly available, and members of another rarely observed class. This often arises when attempting to detect abnormal activity, eg. jet engine failure, computer network intrusions, disease symptoms, etc. In each of these domains, anomalous examples may be scarce or entirely absent during training, but their subsequent identification is of crucial importance.

In this talk I shall describe a new technique for incrementally building a one-class classifier from a sequence of training examples. By contrast, almost all existing one-class classification algorithms require all training examples to be available at once, for a single "batch" learning step: if a new example is presented, the classifier must be retrained from scratch. The choice of an appropriate level model complexity, already a problem when data arrives in a single batch, is even more difficult when data arrives sequentially. The algorithm I propose in this talk outperforms a current state-of-the-art incremental one-class learning algorithm on a variety of datasets, while requiring only an upper limit on model complexity to be specified.

back to seminars

Matthew Szenher,Visual Homing by Maximising Mutual Image Information

Visual homing is the process of returning to a previously visited goal position S from a nearby position C using only an image I captured at S and the current image J at C. Visual homing, unlike many navigation algorithms, makes no attempt to explicitly localise the navigating agent and so requires no landmark map of its environment. In many real-world environments, the root-mean-square (RMS) between I and J increases fairly smoothly as the Euclidean distance between S and C increases. An autonomous robot can use this phenomenon to return to S from C by moving so as to maximise the RMS signal.

Unfortunately, RMS is affected by changes in lighting and landmark location and by proximate landmarks. We describe an image similarity measure -- mutual information -- which is more robust in such dynamic environments. We shall discuss experiments performed using images from a laboratory environment which demonstrate the superiority of mutual information as a similarity measure for use in robotic homing.

back to seminars

Matthew Whitaker,Investigating the effects of inter-agent communication on multi-agent reinforcement learning

This proposal aims to explore the effects of communicated information in a multi-agent system containing reinforcement learning agents. The inter-agent communication of four types of information will be examined - state vectors, reward values, chosen actions and policies. These four types of information are fundamental to all reinforcement learning algorithms. By sharing this information agents can potentially increase their knowledge of the world state, ascertain the goals of other agents in the environment and coordinate their actions to achieve common goals. It is hoped that an examination of how communication influences learning will shed light on the relatively unexplored problem of integrating communication with learning. This talk will discuss the motivations for combining communication and learning, together with a brief overview of existing multi-agent reinforcement learning algorithms. The research questions that form the basis of the proposal will be detailed and some experiments intended to answer them described.

back to seminars

Timothy Hospedales,An Adaptive Machine Director

We model the class of problem faced by a video broadcast director, who must act as an active perception agent to select a view of interest to a human from a range of possibilities. Real time learning of a broadcast direction policy is achieved by efficient online learning of the model's parameters based on intermittent user feedback. In contrast to existing machine direction systems, which are dedicated to a particular scenario, our approach allows flexible learning of direction policies for novel domains or for viewer-specific preferences. We illustrate the flexibility of our approach by applying our model to a selection of scenarios with audio-visual input including teleconferencing, meetings and dance entertainment.

back to seminars

Theophile Gonos,Adaptation to the environment: The role of sensory neurons in sensor self-calibration and response to failure

Robots and animals are agents surrounded by their environment. They sense and act on this environment which is partially observable, stochastic, dynamic and continuous. From the agent's point of view, the environment appears through its sensors. To be relevant, the information that is sent by sensors should perfectly fit the agent's environment. Therefore, the correct calibration of them is essential.

In nature, agents cannot count on supervisors and thus need to self-calibrate. This issue is a major problem in engineering. In this talk, I show how recent studies in biology lead to new ways to solve self-calibration. I then present a perception adaptation mechanism and show how the fit of sensors to environmental changes and their response to failure are linked. I conclude with a discussion of how self-calibration can affect an agent's behaviour.

back to seminars

Toby Collins,Registering and Completing Deformable Isometric Surfaces (or what I did in France)

Having recently come back from an 8 week internship in the University of Clermont Ferrand, France, the focus of this talk will be to give a flavour of the research I have been involved with throughout my time in their labs. The particular research area concerns 3D object registration/model fitting and surface completion, which as a result of ever more pervasive 3D sensor technology is becoming rather hotly persued in the computer vision and graphics communities. Deformable object registration involves the process of aligning a given model to different deformed instances of that object (i.e. establishing dense surface correspondence between the model and the observational data.) This finds considerable application in computer vision tasks (such as object tracking, motion analysis and deformable object recognition) and computer graphics (such as data-driven animation, motion transfer, texture extraction and substitution).

In this talk I will present an approach for registering deforming surfaces in ranges images which, contrary to stat-of-the-art methods does not rely on a matching template or model to be known a priori. For this we consider the class of isometrically deforming surfaces (i.e. those which undergo inextensible deformations such as cloth and paper) where geodesic invariants can be exploited to make the problem sufficiently tractable. The second problem we tackle is surface completion, which arises when we are given only partial surface observations from which to form a complete surface representation. This often occurs when the surface undergoes occlusion from itself or other objects (a flying flag is a good example of this). Again, using properties derived from surface isometric, we have developed the new idea of embedded mosiacing, a process similar to image mosiacing (or image stitching) where a surface composite can be formed by stitching together segments in the surface's intrinsic coordinate spaces rather than the full 3D space. These ideas have been brought together in our BMVC 2007 submission which is currently under review.

back to seminars

Finlay Stewart,Visuomotor behaviour of flying Drosophila

Despite having tiny brains and poor visual acuity, fruit flies are able to negotiate complex 3D environments at high speed. In order to investigate this ability, I have been tracking flies' movement in a free-flight arena. By changing the patterns on the walls, the nature of the fly's visual reflexes can be investigated. Flies typically fly in relatively straight lines interspersed with rapid yaw turns termed saccades. Thus a question of particular interest is the role that the visual environment plays in initiating saccades. A related issue is the degree to which saccades represent fixed motor patterns. I have also found that flies are able to avoid obstacles using smooth, non-saccadic turns, and that this behaviour depends strongly on the visual surroundings. Finally, I shall discuss my attempts to model these phenomena, and some of the particular difficulties I have encountered.

back to seminars

Jeremy Wyatt (School of Computer Science, University of Birmingham),Talking with Robots: A Case Study in Architectures for Cognitive Robotics

In what ways can we integrate multiple types of sensing and action in a robot? This question gets to the heart of deep issues in AI such as the nature and use of representations, and the control of the flow of information in a cognitive architecture. In this talk I will describe some work we are doing on architectures for cognitive robots. I will start by detailing our early work on systems that used and understood utterances about a scene with objects. These experiences led us to design an architecture toolkit for developing cognitive robots. I will describe this, the requirements it is designed to satisfy, its relation to some other architectures, and two early robot systems we have constructed using it.

back to seminars

Francisco Perales (University of Balearic Islands, Spain),Perceptual User Interfaces: new paradigms, some applications and examples

The research of new human-computer interfaces has become a growing field in computer science, which aims to attain the development of more natural, intuitive, unobtrusive and efficient interfaces. This objective resulted in the concept of Perceptual User Interfaces (PUIs) that are turning out to be very popular as they seek to make the user interface more natural and compelling by taking advantage of the ways in which people naturally interact with each other and the world. PUIs can use speech and sound recognition and generation, computer vision, graphical animation and visualization, language understanding, touch-based sensing and feedback (haptics), learning, user modeling and dialog management.

These new interfaces can be used in different scenarios (cars, homes...), but a more important issue is the systems' potential users. PUIs offer assistive technology for people with physical disabilities, which can help them to lead more independent lives and to any kind of audience they contribute to new and more powerful interaction experiences.

Of all the communication channels that interface information can travel through, computer vision provides a lot of information, which can be used for detection and recognition of human actions and gestures that can be analyzed for interaction purposes.

When sitting in front of a computer and using webcams (very common devices nowadays), heads and faces are presumably visible. Therefore, the system is based in head or face feature detection and tracking, and face gesture or expression recognition can become very effective human-computer interfaces. Of course, difficulties can arise from in-plane (tilted head, upside down) and out-of-plane (frontal view, side view) rotations of the head, facial hair, glasses, lighting variations and cluttered background. Besides, when using standard USB webcams image resolution can be very poor - this has to be taken into account.

Different approaches have been used for non invasive face/head-based interfaces. For the control of the position some systems analyze facial cues such as colour distributions, head geometry or motion. Another works track facial features or gaze including infrared lighting. To simulate the user's events it is possible to use facial gesture recognition. In this paper as facial gestures we consider the atomic facial feature motions such as eye blinking, winks or mouth's opening. Other systems contemplate the head gesture recognition that implies overall head motions or facial expression recognition that combines changes of the mentioned facial features to express an emotion.

In this talk, we present an introduction to the main ideas and recent works of PUI and in particular the work done by Computer Graphics & Vision Group at UIB. Some videos and demos will be presented in an informal way.

back to seminars

Michael Mangan,Long range navigation of Desert ants

Desert ants are thermophilic scavengers that forage over large distances for food during the hottest hours of the day. Surface temperatures during these foraging excursions typically range from 40-70 degrees celsius. Thus environmental pressures ensure that desert ants employ efficient navigation strategies when returning to the nest and/or a regular feeding site. These navigational tools are of interest to roboticists and biologists alike therefore desert ants present an excellent model system for Biorobotics studies. In this presentation I will provide an overview of the experimental studies of desert ant navigation and introduce the theoretical models that have been proposed. I shall also discuss the behavioural experiments and robot models which I intend to implement in my PhD to unearth some of the underlying mechanisms of desert ant navigation.

back to seminars

Aroosha Laghaee,Behaviour-based control of an 8 DoF robot arm

I am working on an 8 degrees of freedom (DoF) robot arm for use in a genomics laboratory for pick & placing of microtitre plates and solutions between automated machines. I am using behaviour-based and biologically inspired methods to simplify what would otherwise (by classical methods) be a computationally daunting control problem. I avoid the use of inverse kinematics and the need for precisely modellable engineering by approaching destination positions by iterative sensed approximation to the desired spatial relationships. Adaptability to uncertainty and changes is built into this approach.

In this talk I will describe my general approach. The robot workspace is divided into sections in which a central standard pose of the arm permits a good range of reach using simple movements to which simple linear approximations work well. Larger movements move via the standard poses to the nearest standard pose to the required destination position, after which iterative approximation of simple moves to the destination get as near as is required.

back to seminars

Pak Ho Shum,Interaction-graph: Precomputed Avatar Interactions for Real-time Animation

In this talk, we introduce a new method to generate a realistic animation of avatars densely interacting with each other in a competitive environment. Fighting is used as an example. We propose a new algorithm called the temporal expansion approach which maps the continuous time action plan to a discrete space such that turn-based evaluation methods can be used. In order to avoid the exponential cost of expanding the gametree in run-time, we find the meaningful coupling of the actions by the two avatars and generate a finite state machine (FSM) called the Interaction Graph in the pre-processing stage. The Interaction Graph allows us to control the avatars intelligently in real-time. It is also possible to interactively change the evaluation function to simulate different styles of fights.

back to seminars

Jafreezal Jaafar,A Fuzzy Action Selection Method for Autonomous Navigation in Unknown Environments.

I will present an action selection method using fuzzy logic. The objective is to solve behaviour conflict in behaviour-based architectures for autonomous navigation in unknown environments. Two main problems have been identified: how to decide which behaviour should be activated at each instant; and how to combine the results from different behaviours into one action. The method uses fuzzy α-levels to compute behaviour weight for each behaviour and the final action is selected using the Huwicz criterion. The results show the mapping of inputs to output with a near-optimum path in every navigation task.

back to seminars

Georgios Petkos,Learning dynamics and control under varying contexts: from discrete to continuous contexts

High performance control algorithms require an accurate model of the system's dynamics. Learning dynamics is an attractive alternative when accurate analytical derivation is not possible. However, dynamics may exhibit nonstationarity due to interaction with different environments or objects. Under nonstationary dynamics, conventional learning methods would require unlearning what has been previously learned and adapting to the new environment. If the dynamics switch back and forth, readapting everytime is an inefficient and suboptimal strategy.

In this talk I will discuss a probabilistic setup that avoids the problem of forgetting what has been previously learned. We start by conditioning the dynamics on an unobserved discrete variable. This corresponds to a multiple model scenario, where each model is appropriate to a specific environment. We then replace the discrete variable with a set of unobserved continuous variables. I will show how this model can be obtained in the scenario where nonstationarity results from manipulation of different objects by a manipulator.

back to seminars

Mark Payne,Simple Tricks for Sensory Integration in the Cricket

Female crickets walking in an illuminated arena will move towards a male calling song in a series of straight runs and turns. In common with many other insects they will also respond strongly to rotation of their visual field by turning to stabilise themselves with respect to it. This presents the same paradox treated by Erich von Holst in his classic work of the late 1940s: If each intentional movement generates visual flow, which ought to cause a reaction in the opposite direction, then how does the animal ever turn? Holst's explanation of the paradox relies on the concept of an efference copy, that is a copy of the intentional motor command internal to the nervous system. Such a signal, possibly with some further processing, could serve to predict the visual flow that an animal will receive as the result of its own movement, allowing the nervous system to distinguish the self-generated component of its sensory input from the external.

Unfortunately, predicting the reafferent signal, even for horizontal optical flow, is difficult. For rotation on the spot the signal will be dependent not only on the speed of the movement, but on the spatial frequency, illumination and contrast of the horizontal visual environment. When the rotation is not on the spot, which is often, it gets even harder; the flow depends on the distance of all the objects in the scene and the problem of predicting the reafference looks arbitrarily complicated.

If the view put forward by Wehner (1987) is correct, i.e. "animals solve their problems of spatial navigation not by resorting to abstract computations ... but by adopting approximations, shortcuts and simple tricks" then the cricket ought to be able to teach us some tricks about sensory integration. In my talk I will review the kinds of tricks I have been considering, and my efforts to trick the cricket into revealing them in the laboratory.

back to seminars

Hugo Rosano,Biologically Inspired Compliant Locomotion

Based on neurophysiological and behavioural experiments on the stick insect, it has been suggested that each leg is loosely coupled with others. Additionally, each thoracic segment is known to respond differently to similar stimuli. Nonetheless, all legs move coordinately in a way that almost all body trajectory are possible.

Proposed solutions are based on positive feedback controllers, which in part explain the close kinematic chain problem. However, these have been mostly tested in kinematic models and when implemented in a dynamic model there have been few positive results.

In this talk, I will show what was missing in current walking controller models and then I will present a model that replicates insect manoeuvrability by using compliant locomotion.

back to seminars

Adrian Haith,Implications of non-stationary kinematics and dynamics in motor learning

Most models of motor adaptation do not explicitly distinguish between adaptation to changes in the dynamics of the plant and adaptation to changes in the kinematics of the task. I argue that this distinction must be made explicit since different adaptation strategies are required in each case. I will describe the implications of changing kinematics and dynamics in two contexts: the vestibulo-ocular reflex (VOR) and reaching movements.

In particular, I will describe how the failure of existing VOR adaptation models to distinguish between kinematics and dynamics can lead to slow convergence and instability. In the context of reaching movements, experimental findings on adaptation to kinematics vs dynamics transformations have proven difficult to interpret within existing theoretical frameworks. I will outline plans to extend these frameworks accordingly into a more complete model of human motor adaptation.

back to seminars

Tom Larkworthy,Control for self reconfiguring robots

I am looking at three problems relating to self reconfiguring robots:

• Localization: How can internal components determine their position in a global frame of reference.
• Reactive reconfiguration: How to reconfigure the robot to suit the environment, such as narrowing the morphology to travel down a narrow passage.
• Integrating heterogeneous robotic "tool" components into the structure: How can the specialist tool control algorithms be integrated into the control architecture transparently and in a distributed manner. How can the arbitrary tool morphology be accounted for during reconfiguration.

I will be presenting my robot design and my progress on these three themes.

back to seminars

Tim Lukins,Looking at Construction Progress with Computer Vision

Modern large scale construction projects are incredibly complex to manage and maintain on schedule/budget. Part of the problem is that feedback from surveys and progress assessments are all too infrequent and out-of-date. Recent advances in Building Information Systems provide the framework for integration of ongoing project data, but still require manual input and interpretation. And yet, much hinges on the outcomes in terms of productivity and (ultimately) costs.

So, can we automate this process? Can we get a computer to recognise completed work? Actual construction sites represent quite possibly the worst kind of scene to try and interpret with Computer Vision. However, the ease and frequency with which photographs and video can be taken make visual input an appealing source...

In this talk I present some more of the background to this problem, the challenges, and the applied research currently being undertaken at Heriot-Watt university. This attempts to broadly combine the benefits of high-level geometric object matching for localisation, together with low-level pixel based change detection and texture analysis for status. Some initial results will be shown along with discussion about future directions.

back to seminars

Douglas Aberdeen,Developing RL Approaches for Realistic Road Traffic Control

Road traffic control (RTC) is a decentralised, noisy, and partially observable control problem. Improving real-world traffic control systems, even by a small fraction, would have measurable benefits in large cities. Despite the fact that RTC can obviously be modelled as an RL problem there is little work in the area; perhaps because of the difficulty of solving large decentralised POMDPs. This talk will describe two recent advances in RL and how we have used them to tackle decentralised POMDPs. The first is the use of the Natural Actor-Critic algorithm. The second is the use of Conditional Random Fields to solve the problem of selecting the optimal joint action in a decentralised POMDP --- which is usually considered NEXP-hard in the number of agents. We demonstrate the combination of these two approaches using a sophisticated micro-level traffic simulation that is used to model parts of Sydney's traffic.

Biography

Doug Aberdeen is currently a senior researcher with NICTA, Australia. He has been working in the field of RL for 6 years (including PhD studies), applying RL to many difficult problems such as probabilistic temporal planning. He has recently become involved in road traffic control with the Sydney Road Traffic Authority. He also has interests in AI planning and cluster computing.

back to seminars

Matthew Howard,Learning Utility Surfaces for Movement Selection

Kinematically and dynamically controlled redundant manipulators are commonly controlled through optimisation with respect to some cost function. The form of that function (usually hand-specified by the agent designer) determines the unique preferred action among the infinite set of possible alternatives. In my talk I will describe my work on the alternative approach of extracting and modelling the cost function from motion data, and using it to transfer behaviour from one system to another.

back to seminars

Shu Lim Ho,Extracting topological relationships between avatars with close interactions

The aim of my current research is to develop a topological representaion between the body segments of virtual characters during close interactions such as wrestling and dancing. The new representation will be based on the tangle concept developed in Knot Theory. The pattern of the entangled body segments can be recognized and represented by an invariant called Continued Fraction. Our concept can be used as an index for contents based retrieval of general 3D objects which are entangled, and also as a map for path planning of humanoid robots who are conducting close contact movements with humans.

back to seminars

Tom Erez,Gait Generation using Manifold Control

In this talk, we present a solution for a class of high-dimensional, non-linear dynamical systems where the solution is known to be periodic, as in the case of gait generation. In such systems, a good solution forms a limit cycle in state space. We describe a method for solving such problems using reinforcement learning, and present its application to a highly nonlinear problem, with 14-dimensional state and 4-dimensional action spaces. By focusing the computational effort only where it is most needed - along the one dimensional manifold spanned by the limit cycle - this approach allows us to escape the curse of dimensionality associated with the full state space. Our algorithm applies a novel algorithm for direct policy gradient estimation due to Munos, and although the learning is local, we learn both the desired trajectory AND the policy that realizes it, and so avoid restricting our solution space. My main results are presented in these two movies: MOVIE 1, MOVIE 2.

Time permitting, I would also discuss a related project of learning to walk on uneven terrain This 4-dimensional system, the "copass gait walker", is capable of passive walking over a small range of slopes. We added actuation at the hip joint, and learned a local policy over a one dimensional manifold approximating the limit cycle. By following a shaping protocol, we greatly exteneded the range of slopes the walker can safely traverse. Then, we united several such one-dimensional manifolds to create a composite policy that can handle rough terrain.

Biography

Tom Erez is currently a PhD student at Wahington University in St. Louis, working with William D. Smart on reinforcement learning in continuous domains. Tom studied Mathematics at the Hebrew University of Jerusalem, and spent the previous two years as a researcher in Turin, Italy as part of the EU project GIACS.

back to seminars

Guglielmo Tamburrini,Learning automata and responsibility

Epistemic limitations concerning prediction and explanation of the behaviour of machines that learn from experience are selectively examined by reference to machine learning methods and computational theories of inductive learning. Moral responsibility, liability ascription policies, and other applied ethics problems concerning learning machine operations are discussed in the light of these epistemic limitations.

back to seminars

Tomoyuki Nishita,Rendering Methods for Various Natural Phenomena

The simulation of various natural phenomena is one of the important research fields in computer graphics. In particular, aspects such as sky, clouds, water, fire, trees, smoke, terra ins, desert scenes, snow and fog are indispensable for creating realistic images of natural scenes, flight simulators and so on. Therefore, a lot of researchers have been trying to develop methods for simulating and rendering these. In my presentation I focus on sky, clouds, smoke, desert scenes and atmospheric effects, such as shafts of light. These phenomena have the common feature that they are consist of the effects of small particles. To create realistic images, physical based simulation and rendering are required. In particular, the color greatly depends on the properties of light scattering due to particles. I would like to introduce efficient methods for creating realistic images of such natural phenomena.

Biography

Prof. Nishita received Research Award on Computer Graphics from Information Processing Society of Japan in 1987, and also received Steaven A. Coons award from SIGGRAPH in 2005. He has written twelve SIGGRAPH papers. He is one of the pioneers of Radiosity Method. He was a member of the Editorial board of the IEEE Transactions on Visualization and Computer Graphics. He has lectured at The University of Tokyo since 1994.

back to seminars

Narayanan Edakunni,Fast Online Regression using a Randomly Varying Coefficient Model

In this talk I would be presenting a probabilistic formulation for function approximation using locally linear models. The Randomly Varying Coefficient(RVC) model as it is called learns online and adapts its complexity based on the training data. The use of Variational Bayesian EM yields efficient and online updates for the parameters and a product of experts kind of combination of local models yields smooth estimates for the function that we aim to approximate. Furthermore I will also discuss its salient features over other corresponding learners especially its improved computational efficiency which makes it suitable for real time online learning.

back to seminars

Scott Blunsden,What are they doing?

This talk will focus upon classifying interactions between people when viewed through a camera. Previous research has focused upon describing and classifying individual actions (such examples could include walking or running). Such an approach can lead to shortcomings when trying to describe situations where more than one person is involved. An example of this would be the case of people meeting where it is necessary to take into account both parties to understand the individual actions.

In this talk I shall describe some work we have been doing on how to tackle this problem and show different approaches and results for classifying interactions between multiple people.

back to seminars

Sebastian Bitzer,Dimensionality Reduction for Movement Data

I will introduce a project linking dimensionality reduction and learning control which I am currently preparing. Ijspeert et al. have proposed a way of controlling robots by using control policies based on simple differential equations. This approach has the advantage that robot movements become robust against perturbations and similar movements can be produced by simple scaling of an existing control policy. But most importantly, these control policies can easily be learned from example trajectories. The learning, however, gets increasingly difficult with the dimensionality of the trajectories, i.e. the number of joints involved in the movement. The project will look into applying dimensionality reduction to movement data to make learning easier while keeping advantageous properties of the control algorithm. I will present two suitable candidate methods for nonlinear dimensionality reduction for movement data: Gaussian Process Latent Variable Models and Laplacian Eigenmaps Latent Variable Models.

back to seminars

Leslie Smith(Stirling University),Onsets in sound: what they are and why they are interesting?

There is a large difference between the physical characteristics of sound and what we hear. One aspect of this is our particular sensitivity to sudden increases in energy in the sound. These onsets are characterised by their time of occurrence, their spectral location, their duration, and the rate of increase in the different areas of the spectrum. A biologically inspired onset detector has been developed, and used for sound direction finding in noise. The detector, and current and future directions for this type of work, including possible application to speech will be discussed.

back to seminars

Jan Wessnitzer,Associative learning in the Mushroom Body neuropil of the insect brain

The Mushroom Body is a prominent region for multimodal sensory integration and associative learning in insects. Using realistic neural dynamics and a biologically-based learning rule (spike timing dependent plasticity), the developed model is tested as part of an insect brain inspired architecture within closed loop behavioural tasks.

We could show that the distinctive neuroarchitecture is suited for pattern recognition and is sufficient for non-elemental learning. In a second task, replicating in simulation an experiment carried out on bushcrickets, we show the system can successfully associate visual to auditory cues, so as to maintain a steady heading towards an intermittent sound source.

I will conclude with a discussion on directions for future work.

back to seminars

Barbara Webb,Protocognition

Traditionally, `cognitive' has been used to describe systems that deal with knowledge; that can reason and plan using internal representations. Recently, some researchers have been keen to apply the term `cognition' to any system capable of adaptive interaction with its environment, but this seems too liberal. Clearly, there is a large middle ground of systems with higher capabilities than reactive behaviour but not capable of explicit knowledge representation. In this seminar I will explore some of this middle ground, using examples mostly drawn from insect biology, to open a debate on what minimal or critical capabilities are the precursors to cognition.

back to seminars

Ernesto Andrade,Event Detection Models for Crowded Scenes

This talk will present work on the evaluation of an automatic technique for detection of abnormal events in crowds. Crowd behaviour is difficult to predict and might not be easily semantically translated. Moreover it is difficult to track individuals in the crowd using state of the art tracking algorithms. Therefore we characterise crowd behaviour by observing the crowd optical flow and use unsupervised feature extraction to encode normal crowd behaviour. The unsupervised feature extraction applies spectral clustering to find the optimal number of models to represent normal motion patterns. The motion models are HMMs to cope with the variable number of motion samples that might be present in each observation window. The results on simulated crowds analyse the robustness of the approach for detecting crowd emergency scenarios observing the crowd at local and global levels. The results on normal real data show the effectiveness in modelling the more diverse behaviour present in normal crowds. These results improve our previous work in the detection of anomalies in pedestrian data.

back to seminars

Hannah Dee(Leeds University),Where are they going? Spatial cognition and surveillance

Research in computer vision often aims to model patterns of human or animal behaviour - particularly in the surveillance domain. Most of this work derives typical patterns of motion within a scene, or typical interactions, whilst ignoring the psychological processes underpinning the observed motions or interactions. Philosophers and historians of science talk of "levels of explanation" - physical, biological, design, intentional... and there is a real sense in which the computer vision communities approach thus far has been largely at the physical level of explanation.

This talk will describe work showing that the incorporation of higher level intentional reasoning can help to abstract away from specific environments and can enable easily interpretable summarisations of human behaviour. This enables us to talk of an agent's goals, and to explain the observed behaviour rather than merely to describe it.

back to seminars

Heiko Hoffmann,Tactile input improves robot-arm control under changing context

Adaptive robotic motor control under changing hidden context (e.g., mass and size of a manipulated object) is an active research area that lacks a satisfactory solution. I will present a realistic robot-arm simulation that shows that tactile sensors improve motor control under hidden context. In the special case of a single continuous context variable (here, mass), the use of tactile-sensor values instead of control torques allows to infer more accurately the unknown mass of the object hold in the robot's hand. This mass in turn can be used to predict control torques. For the more general case of multiple hidden context variables, I demonstrate that a mapping from robot-state and sensor values onto torques can be used for accurate control; the above inference of mass fails in this case.

back to seminars

Sethu Vijayakumar,Optimality principles in sensorimotor control

In this talk, I will look at optimality principles in sensing and motor actions and describe a unified paradigm for control under uncertainty and noise. While traditional emphasis has been on optimizing desired movement trajectories while ignoring sensory feedback, recent work has redefined optimality in terms of feedback control laws I will describe the optimal cue integration and optimal state estimation paradigms, while pointing to work going on in our group related to this topic. Then, I will briefly explain the concept of stochastic optimal control and discuss differences between open-loop and closed loop optimization: while open loop estimation with feedforward commands and feedback to correct for noise and deviation has been the norm, closed loop optimization may be the right way to go, albeit still being relatively less well understood for non-toy, real world problems.

back to seminars

Prof. Shigeo Morishima,Instant Movie Cast Generator in Future Cast System

"I wish I were a movie star"? This is a dream most of us share. Now, the "Future Cast" system makes this dream a reality. Future Cast is an unprecedented viewer-participation-based visual attraction that provides viewers with an amazing experience impossible with conventional amusement facilities: the chance for each and every member of the audience to actively take part in the movie.

Future Cast premiered at the MITSUI-TOSHIBA Pavilion during EXPO 2005 (Exposition of Global Harmony) in Aichi, Japan. Future Cast attracted a huge number of visitors and became one of the EXPO's most popular attractions with as many as 1.63 million visitors enjoying this "next-generation movie" experience during the six-month-long EXPO. Anyone can participate in Future Cast after standing in front of a 3-D scanner for a few seconds to have their image scanned with a simple and completely safe process. The computer analyzes and processes the scanned data to creating a facial image fully automatically. The Future Cast system then determines the gender and age of each "actor" and assigns them an appropriate role to play in the movie. Within just a few minutes the data is converted into personal actor model and the actor begins to speak and perform in a fully CG-based movie as vividly as any real actor. In my talk, I will introduce a making and background story of this Future Cast, and show you also a future vision included in our project "Dive Into the Movie".

Biography

Dr. Shigeo Morishima was born in Japan on August 20, 1959. He received the B.S., M.S. and Ph.D. degrees, all in Electrical Engineering from the University of Tokyo, Tokyo, Japan, in 1982, 1984, and 1987, respectively. From 1987 to 2001, he was an associate professor and from 2001 to 2004, a professor of Seikei University, Tokyo. Currently, he is a professor of School of Science and Engineering, Waseda University.

His research interests include Computer Graphics, Computer Vision, Multimodal Signal Processing and Human Computer Interaction. Dr. Morishima is a member of the IEEE, ACM SIGGRAPH and the Institute of Electronics, Information and Communication Engineers Japan(IEICE-J). He is a leader of Human Communication Science Special Interest Group in IEICE-J and a trustee of Japanese Academy of Facial Studies. He received the NICOGRAPH paper awards in 1988, 1990, 1993, 1996 and 2000. Also he received the IEICE-J achievement award in May, 1992 and the Interaction 2001 best paper award from the Information Processing Society of Japan in Feb. 2001. Dr. Morishima was having a stay at Visual Modeling Group, Department of Computer Science, University of Toronto from 1994 to 1995 as a visiting professor. He is now a temporary lecturer of Meiji University and Seikei University, Japan. Also he is a visiting researcher of ATR Spoken Language Communication Laboratory from 2001.

back to seminars

Eric McKenzie,Towards a new method for Level of Detail in Computer Graphics

This talk will describe some work that has been trickling on through the last few years. Level of Detail research at Edinburgh goes back around 10 years. I will briefly review that work to set the background for the current research. It was thought that scale-dependent curvature flow algorithms could provide a new way to control level of detail display. I will show why that doesn't work and how that investigation led to a more positive line of research. The first part of the talk repeats what a few people heard last time with an update to what has happened since.

Level of detail display involves selection of an appropriate resolution at which to display an object in a scene according to its position and movement within the scene.

back to seminars

Taku Komura,Simulating the interactions of multiple characters

Recently, there have been a lot of researches to synthesize / edit the motion of a single avatar in the virtual environment. However, there has been not much work in simulating continuous dense interactions of multiple avatars. In this talk, I will explain about two on-going projects to handle such situations. The first project is about generating a realistic fighting scene based on motion capture data.

We propose a new algorithm called the temporal expansion approach which maps the continuous time action plan to a discrete causality space such that turn-based evaluation methods can be used. Using our method, avatars will plan their strategies taking into account the reaction of the opponent.

The second project is about representing the status of the two characters by using topological information. This approach is effective for motion synthesis, path-planning and contents based retrieval for motions of more than two avatars, such as dancing, wrestling, and wearing / taking off clothes. Preliminary results of the two projects will be presented.

back to seminars

Leslie Pack Kaelbling, Life-Sized Learning

In the last 10 years, the combination of techniques from machine learning and operations research has allowed major advances in learning and planning for uncertain environments. Reasonably large problems can be solved using current techniques. But what if we want to scale up to the uncertain learning and planning problem that you face every day? It is many orders of magnitude larger than the biggest problem we can solve currently.

In this talk, I'll describe three pieces of work that try to begin to address working in truly huge environments. The first is a method for learning probabilistic rules to describe naive physics models of the interactions between objects. The second is an uncertain planning algorithm that uses the rules learned by the first method to construct contingency plans that consider enough cases to perform robustly, but are much smaller than complete policies. The last piece is preliminary work on combining multiple abstraction methods dynamically, in order to allow an agent to have a working model of the environment that changes focus depending on the current situation.

back to seminars

Bob Fisher, Many randomly distributed single pixel cameras / An Empirical Model for Saturation and Capacity in Classifier Spaces

This talk presents two recent pieces of research:

1) Scene recovery problems when interpreting data from massive numbers of randomly distributed single pixel cameras. Assuming that the camera positionsand orientations are approximately known, the talk shows that both distant and nearby scenes can be reconstructed, and analyzes how recovery performance varies with sensor parameters.

2) When assessing reported classification results based on selection of members from a database (e.g. a face database), one would like to know what is an achievable classification rate, given the noise level, dimensionality of the feature set and number of classes in the database. As best we can tell, no general results exist for this question, although many classification rates appear in different papers. This talk presents an empirical formula for MAP classification that links the number of discriminable classes to the error rate, dimensionality of the feature data and the feature noise level.

back to seminars

Matt Howard, Inferring Utility Surfaces for Movement Selection

The aim of this Ph.D project is to research and develop algorithms for inferring the cost functions used in human or humanoid motor control directly from observations of motion.

Humans and humanoid robots are by nature highly redundant systems with respect to the typical tasks they are asked to perform. This redundancy manifests itself in the large numbers of control degrees of freedom far exceeding that required for most tasks. The result is a large space of possible movements, any one of which may be chosen to fullfil the task constraints. An intuitive model for determining which movement is chosen is to define a cost function over possible movements, such that the movement chosen is deemed optimal in some way. Such functions define a unique preferred strategy for movement selection and define a metric by which movements can be optimised.

However, the definition of such metrics is not a trivial matter, and presents problems both in the design of behaviour (where the task is to determine a quantitative metric that captures some qualitative notion of 'desired' behaviour) and the analysis of behaviour (where the task is to determine a metric whose optimisation explains stereotypical features of the behaviour). So far in both cases the problem has been approached by a 'proposition-evaluation' strategy where some (hand-designed) cost function is first proposed and is then evaluated against the observed/desired motion.

In this talk I will propose a different approach whereby cost functions are directly modelled in a data-driven way from observations of task-oriented movements. I will present work where an instantaneous cost function is extracted from movement data of a constrained, kinematically-controlled manipulator and propose extensions of this work to deriving cost functions under time-varying constraints. Furthermore, by drawing analogies with recent work in Inverse Reinforcement Learning I will suggest how time-integral cost functions can similarly be derived.

back to seminars

Paolo Favaro, Inference of shape and radiance for modelling and Interaction

Images are measurements of light emitted from a scene, which in general depend on the shape, colour, and motion of the objects in the scene, and on the surrounding illumination. While in computer graphics one is typically interested in simulating images given the scene, in computer vision one is interested in the inverse problem, where the images are given, and one seeks the shape, colour and motion of objects and the illumination of the scene. In this presentation I will give an overview of the various inverse problems that I have been working on. In particular, I will briefly present work on: 1. Reflectance, illumination and 3D surface estimation in multiview geometry 2. Image restoration and 3D shape estimation in the presence of defocus and motion blur 3. Dynamic texture segmentation 4. Real-time structure from motion 5. Detection and classification via adaboost

back to seminars

Masashi Sugiyama, Value Function Approximation on Non-Linear Manifolds in Reinforcement Learning

The least-squares approach works efficiently in value function approximation, given appropriate basis functions. Because of its smoothness, the Gaussian kernel is a popular and useful choice as a basis function. However, it does not allow for discontinuities which typically arises in real-world reinforcement learning tasks. In this talk, I introduce a recently proposed basis function based on geodesic Gaussian kernels, which exploits the non-linear, discontinuous manifold structure induced by the Markov decision processes. The usefulness of the proposed method is successfully demonstrated in a simulated robot arm control and Khepera robot navigation.

This is a joint work with Hirotaka Hachiya, Christopher Towell, and Sethu Vijayakumar supported by the EuMI Programme of European Commission.

back to seminars

Prof. David Lane,Autonomous Underwater Vehicle Research in the Ocean Systems Laboratory

The talk will overview some active research themes and programmes on Autonomous Underwater Vehicles in the Ocean Systems Laboratory at Heriot-Watt. Themes in SLAM Navigation. servoing/tracking,.mission planning, computer aided detection/classification and biosonar design should feature. Experiences productising these capabilities and turning them into whole product solutions addressing the needs of markets will be included. Opportunities for involvement through the annual SAUC-E European Student Autonomous Underwater Vehicle Challenge will conclude.

back to seminars

---

Institute of Perception, Action, and Behaviour © 2006
(*) Images show the Honda Asimo robot, DLR-LWR arm, Sony AIBO, Koala and Khepera robots.