Tin men in pinnies

Robotics is far too difficult to be left solely to engineers and computer scientists. If we are to succeed in getting robots to do menial jobs for us, we need to understand the principles behind intelligent behaviour of autonomous systems, and animals provide the obvious, indeed the only model. If you accept this, then Intelligent Behaviour in Animals and Robots looks incredibly exciting. If I tell you also, that by chapter 10 the authors have specified a design for a robot that will wash and put away the dishes, and clean the floors and windows, then your appetite should be well and truly whetted. Unfortunately, however, disappointment awaits.

The theory presented starts from the view that "intelligence" is not a property of a system but of the behaviour of that system in a particular environment. Therefore, the house-cleaning robot need not have sophisticated cognitive mechanisms for producing the appropriate behaviours - all it needs is simple but well-adapted ones for the tasks at hand. This is in stark contrast to the classical artificial intelligence approach, where instead of having one high-level goal such as clean the house, the system has a large number of simple skills that cumulatively achieve the same end. In the former type of system the problem is how to achieve the goal, whereas in the latter the problem is what to do and when. This cannot simply be a matter of time allocation but requires a continuous reassessment of the costs and benefits of performing one behaviour over another.

So how do animals and how should robots do this? The core of the book argues the case for utility theory in which the agent has some internal value system that allows it to assess the costs and benefits of performing one activity over the rest. We are told how to combine individual utility functions into one big goal function and how to tell whether this will lead to stable behaviour or not.

The problems with the book are not so much to do with what it tells us but with what it does not. For example, it does not tell us how to go about decomposing house cleaning into simple enough modules for current robots to perform. The design given for this robot only reduces the problem to timesharing between recharging the robot's battery, washing and putting away the dishes, and cleaning the floors and windows. These individual tasks are still far too difficult to implement in a robot. For example, to work out whether it should be washing up, the robot requires sensors that detect the number of dirty plates, but computer vision is still not sophisticated enough to distinguish plates from non-plates when they are in the real world of the kitchen, let alone distinguish dirty plates from clean ones. Even recharging the battery is problematic because the robot would need to navigate back to its nest and robust navigation is extremely difficult. Therefore, the design of the house-cleaning robot would be much more convincing if the robot could also evaluate the cost of getting lost and the cost of trying to wash paper plates and so on.

If the book had used decompositions of animal behaviour to illustrate how to decompose complex tasks into simple ones then it would have been truly radical. The individual skills of the robot must be very simple indeed and therefore complex behaviour must be made up of very many simple skills. However, this then becomes a problem for decision-making through cost-benefit analysis. Even with four activities the goal function is complicated and time-consuming, and the authors acknowledge that the number of candidate activities at any one time must be small. They propose that the total number of candidate activities is reduced to manageable proportions through some review procedure but they never explain how such a procedure could work.

Of course, ethological examples do not help much when it comes to house cleaning. Which leads to the question of what role the ethological examples play. The authors argue that evolution can be considered as a designing agent and that the design of robots for a competitive marketplace is an evolutionary process but this analogy takes us nowhere. Perhaps it is a veiled insult aimed at roboticists for relying on blind trial and error. In fact, the ethological examples in this book have very little to contribute to robotics at all. It seems as if they have been included merely for atmospheric support: utility theory can describe some aspects of animal behaviour and utility theory can help in some aspects of the design of robots. This is hardly radical especially considering that utility theory has been around for such a long time.

Robotics has plenty of theory but still very little behaviour to show for it. Psychology, on the other hand, has more than enough behaviour but has failed to come up with a unifying theory. Both disciplines are studying autonomous agents, so some sort of marriage would seem ideal. There is no reason why robotics should not inform psychology just as much as the other way round. If psychologists understood the practical difficulties involved in dealing with the world even in the most simple ways, then our understanding of the problems that cognition must solve would improve. The book does not attempt this other direction of information transfer, nor does it provide one example of behaviour from a real robot to support the theory. In fact, the only example of real robot behaviour in the whole book comes on the last page and illustrates the possibilities of emergent functionality. If this had come earlier in the book it would have been better, but I suspect that emergent functionality is not so easy to resolve with the utility theory approach. How could a robot assess the consequences of behaviours with emergent functionality?

Intelligent Behaviour in Animals and Robots is not without merit. It makes many interesting distinctions, such as what it means to be autonomous, and it is well argued and well presented. Its greatest merit is that it is trying to do something that is potentially very exciting. But it could be so much better.

Benedict St Johnston is a research associate in the Robotics Laboratory, University of Edinburgh.