Rule-based systems employed to model complex object behaviours, do not necessarily provide a realistic portrayal of true behaviour. To capture the real characteristics in a specific environment, a better model may be learnt from observation. This paper presents a novel approach to learning long-term spatio-temporal patterns of objects in image sequences, using a neural network paradigm to predict future behaviour. The results demonstrate the application of our approach to the problem of predicting animal behaviour in response to a predator.

The Robot Sheepdog Project has developed a mobile robot that gathers a flock of ducks and manoeuvres them safely to a specified goal position. This is the first example of a robot system that exploits and controls an animal's behaviour to achieve a useful task. A potential-field model of flocking behaviour was constructed and used to invesigate methods for generalised flock control. One possible algorithm is described and demonstrated to work both in simulation and in the real world.

We consider the problem of monitoring and controlling the position of herd animals, and view animals as agents with natural mobility but not strictly controllable. By exploiting knowledge of individual and herd behaviour we would like to apply a vast body of theory in robotics and motion planning to achieving the constrained motion of a herd. In this paper we describe the concept of a virtual fence which applies a stimulus to the animal as a function of its pose with respect to the fenceline. Multiple fence lines can define a region, and the fences can be static or dynamic. The fence algorithm is implemented by a small position-aware computer device worn by the animal, which we refer to as a Smart Collar. We describe a herd-animal simulator, the Smart Collar hardware and algorithms for tracking and controlling animals as well as the results of on-farm experiments with up to 8 Smart Collars. 1

will obtain a networked system that can function as an information backbone for the group. Information can flow across this group to update individual parameters and programs (for example the motion plans for the virtual fences), coordinate tasks, USA, (e-mail: ################################ ). CSIRO Manufacturing & Infrastructure Technology, Australia, (email: ######### ). USA, (e-mail: ### ### ). Computer Science and Artificial Intelligence Laboratory, MIT, Cambridge MA 02139, USA, (e-mail: ################### ). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Copyright 200X ACM X-XXXXX-XX-X/XX/XX ..

In the multi-agent community, the need for social deliberation appears contradictory with the need for reactivity. In this paper, we try to show that we can draw the benefits of both being reactive and being socially organized thanks to what we call social reactivity . In order to defend this claim, we describe a simulation experiment in which several sheepdog agents have to coordinate their effort to drive a flock of ducks towards a goal area. We implement reactive controllers for agents in the Classifier Systems formalism and we compare the performance of purely reactive, solipsistic agents which are coordinated implicitly with the performance of agents using roles. We show that our role-based agents perform better than the solipsistic ones, but because of constraints on the roles of the agents, the solipsistic controllers are more robust and more opportunistic. Then we show that, by exchanging reactively their roles, a process which can be seen as implementing a form of social deliberation, role-based agents finally outperform the solipsistic ones. Since designing by hand the rules for exchanging the roles proved difficult, we conclude by advocating the necessity of tackling the problem of letting the agents learn their own role exchange processes.

Abstract — This paper considers the problem of controlling a group of agents under the constraint that every agent must be given the same control input. This problem is relevant for the control of mobile micro-robots that all receive the same power and control signals through an underlying substrate. Despite this restriction, several examples in simulation demonstrate that it is possible to get a group of micro-robots to perform useful tasks. All of these tasks are derived by thinking about the relationships between robots, rather than about their individual states. I.

This thesis describes the development of an autonomous robot system that gathers a ock of ducks in a circular arena and manoeuvres them safely to a pre-determined goal position. In the process it establishes a methodology for developing robots that interact with animals. An important feature of this methodology is that it enables the development of a machine that can usefully interact with an animal without using the animal in the design process. Interacting with animals imposes strong constraints of real-time action, robustness and animal safety. A suitable arena, robot vehicle, control architecture and vision system are described. An animal-interactive robot must be robust with respect to the inevitable variations in behaviour between individual animals and even in the same animal over time. It is suggested that (a) animal-interactive robot controllers should exploit the underlying mechanisms of the subject animals' behaviour rather than the details of any particular animal or gro...

This paper is about the use of roles in a multiagent adaptive context. We describe a simulation experiment in which several sheepdog agents have to coordinate their effort to drive a flock of ducks towards a goal area. We use the Classifier System formalism to control the agents. We show that using a notion of role is natural in such a context. We compare the performance of an expert controller with and without roles. Then we show how applying adaptive techniques to that bootstrap controller can improve the performance with respect to expert rules. From this empirical study, it appears that an emergent strategy gets better results than the conception of the roles we designed by hand. Thus we advocate the necessity of tackling the problem of evolving the roles of the agents.

In complex simulations involving several interacting agents, the behavior of the overall program is difficult to predict and control. As a consequence, the designers have to adopt a trial-and-error strategy. In this paper we want to show that helping experts to design simulation automata as classi er systems (CSs) by hand and using a semi-automated improvement functionality can be a very e cient engineering approach. Through the example of a simple multiagent simulation, we show how simulation automata can be implemented into the CS formalism. Then we explain how the obtained CS can be improved either by hand or thanks to adaptive algorithms. We first show how giving indications on the non-Markov character of the problems faced by the classifiers can help the experts to improve the controllers and we explain why adding modularity in the CS formalism is important. Then we show how the adaptive algorithms inherent to Learning Classifier Systems (LCSs) can be used in such a context, we discuss...

We describe a moving virtual fence algorithm for herding cows. Each animal in the herd is given a smart collar consisting of a GPS, PDA, wireless networking and a sound amplifier. Using the GPS, the animal's location can be verified relative to the fence boundary. When approaching the perimeter, the animal is presented with a sound stimulus whose effect is to move away. We have developed the virtual fence control algorithm for moving a herd. We present simulation results and data from experiments with 8 cows equipped with smart collars. I.