OUCL Robot Sheepdog Project

Exhibiting at the Royal Society Summer Science Exhibition (3-5 July 2001)

Robot Sheepdog Project

Sheepdog with duck flock in Lancashire, 1996

The original project (1996-1998)

Stephen Cameron
Research student: Richard Vaughan

People have been building robotic systems that work with humans for many years, and they generally assume that the humans are cooperative; but what about robots to work with animals? It was this thought that led to the "Robot Sheepdog Project" (RSP) at the Silsoe Research Institute, which brought together 3 research students based in Oxford, Leeds and Bristol, together with Silsoe's experience at dealing with farm animals.

In fact, it was soon decided that herding sheep was too big a jump in one step, and so the robot herds ducks instead -- but as these are used to train sheepdogs we claim that the results could be applied to sheep! The robot uses a fixed television camera to see what the ducks are doing, and our student takes this information, plus a mathematical model of the ducks' behaviour, and uses this to control the robot.

This may sound straightforward, but nothing like it has ever been tried before; we had to design the robot, guess the correct mathematical model, and implement it whilst ensuring that the ducks were never harmed! To do this we first relied a lot on simulation, and on seeing whether the mathematical model gave simulated results that "looked right" to expert duck herders. Finally, the system was tried with real ducks, and was shown to be capable of herding a flock of ducks from one end of an arena to a specified position at the other end.

Publications

Robot Sheepdog Project achieves automatic flock control, in Proceedings of the International Conference on Simulation of Adaptive Behaviour, Zurich, Switzerland (1998).
Robot control of animal flocks, in Proceedings of the International Symposium on Intelligent Control, Gaithersburg, Maryland, USA (1998).
Experiments in Automatic Flock Control, Journal of Robotics and Autonomous Systems 31:109-117, April 2000, © Elsevier Science BV.

Articles

Robot rival to shepherd's best friend, Electronic Telegraph (7 Aug 1998)
More byte and less stress, BBC News Online (24 Nov 1998)
World first for Rover the Robot, SRI News (Winter 1998)

The current project

Stephen Cameron, Joe Pitt-Francis, Irina Voiculescu
Research students: Shawn Rusaw, Tristam Fenton-May

Descartes robot from Diversified Enterprises (www.divent.com) The Spatial Reasoning research group is continuing the research work, but using robot sheep as well as robot sheepdogs to extend the theories without the complications of dealing with real animals.

The project uses a collection of small wheeled mobile robots, each of which can act as a sheep or a dog. The "sheep" behave according to a pre-programmed flocking model, whereas the "dogs" are controlled through an external PC. The dogs can be driven by a human using a joystick, or can be programmed to herd the sheep to a predetermined point.

The projects objectives are

to develop a scientific understanding of the way in which flocking and herding behaviour in animals occurs;
to investigate how collaboration between robots allows them to perform tasks that they could not perform alone;
to investigate the scientific feasibility of robot "guard dogs".

Graph showing that duck is attracted to duck but repelled by walls and by the dog The system uses an overhead camera that tells a central computer where each animal is. Thus the controller knows about all the animals, but the robots can be kept simple and cheap as they contain few processing or sensing mechanisms. The PC sends low-level commands to the robots via infra-red links.

The model of sheep and duck behaviour is very simple and contains just three terms: attraction to one another, repulsion from walls and other fixed obstacles, repulsion from the dog. Although the model contains only these three terms, plus a random "jitter", there is a realistic tendency for the robotic sheep and ducks to flock.

To move the flock in a particular direction, an individual sheepdog has to move until the flock is between itself and the goal, then gently move forward until the animals are at the goal, and finally back away.

For all the animals, it is interesting to see complex emergent behaviour develop from these simple components. The simplicity of the models which can show realistic behaviour suggests that certain sorts of animal (and human?) behaviour can be quite simply analysed, and raises the possibility that an analysis of human behaviour could, for example, be used to design safer sports stadia.