From Isolated Components to Cognitive Systems
by Christian Balkenius
IKAROS is an infrastructure that will allow large-scale cognitive systems to be built by assimilating and combining results, methods and models from different research groups.
Before robots can move into our homes, offices and hospitals, a certain level of autonomy is needed. We need systems able to function in a human environment without causing too much annoyance - a high level of autonomy, but far less than you would need for, for instance, mission-critical operations in remote or hostile environments.
While it may be acceptable for a service robot to ask us whether we want fillet mignon or a hamburger for dinner, it will not be acceptable for a robot to ask for advice on how to open a door or how to go to the kitchen. In short, it will need the ability to solve all the small everyday problems that humans do not even have to think about. This type of everyday cognition relies as much on sensory processing and motor control as on high-level reasoning and planning. In fact, it uses all the subsystems studied within the field of cognitive science, ranging from emotion and motivation to attention and motor control. An understanding of the biological solutions in these areas may thus give important insights into the architecture that is required.
A common objection to biologically inspired systems is that "aeroplanes do not flap their wings", but it is important to realise that birds and jumbo-jets are solutions to two quite different design problems. When the design problems do coincide, such as in energy-efficient gliding flight, the design solutions are remarkably similar, such as the wing configuration of an albatross and a glider.
We want to suggest that the problem domain for the control systems of future autonomous robots is sufficiently similar to that of the human brain to merit a study of its biological counterpart. A domestic robot will need to exist and function in an environment that is remarkably, though unsurprisingly, well adapted to human cognition and motor behaviour. For humans, there is ideally no design problem; the environment is already adapted to us. For a robot, the best solution is to employ design strategies that are known to work in the environment, and those strategies are those that humans already use.
Today, a large number of computational models are available for many cognitive domains and brain regions involved in cognition. Although many of these models do not correctly reproduce all the functions of their corresponding brain areas, they are often sufficiently developed for use as components in larger systems. An important insight from our earlier research has been that many cognitive phenomena are system properties. The exact operations of the individual components are often not as important as how they are combined into a larger system.
The need for an infrastructure that would allow integration of models and methods from different cognitive research fields and different groups provided the motivation for the IKAROS project (named in tribute of a previous high-flying exercise in biological imitation). The project started in 2001 at Lund University Cognitive Science and builds upon our previous experience with simulators and control systems for robots. The goal is to develop tools that will allow full-scale cognitive architectures for future autonomous robots to be developed based on insights gained from human cognition.
The strength of IKAROS is its ability to assimilate different types of models and algorithms in such a way that they can be used as parts of larger systems. This makes the system open-ended, since support for new methods and hardware can be easily added. This is combined with support for communication between individual modules running on the same or different processors or on different computers on the Internet.
IKAROS systems are also inherently scalable. For example, the same architecture that runs on a small robot with a micro-controller and simple sensors can run on a large cluster computer with a full-scale vision system. Changing the inputs or outputs will automatically scale all individual components in the system to match. Another type of scalability is that architectures built with IKAROS can be moved seamlessly from simulation to real robots.
However, the most important aspect of IKAROS is that it is fully open and freely available. New versions are regularly published on the Internet and it runs on most computer platforms and operating systems. The first public version was released in February 2003. The current distribution already contains modules simulating many different brain regions, modules for image-processing and a collection of biologically motivated learning algorithms. The system also includes modules that communicate with external devices such as video cameras. An interface for mobile robots is currently under development.
We encourage other research groups to use IKAROS for their own research and to contribute new modules to the system. In the future, IKAROS will also include a database of benchmark problems to test and validate different cognitive architectures.
Christian Balkenius and Jan Mørn, Lund University, Cognitive Science, Sweden
Tel: +46 46 222 32 51, +46 46 222 85 88
E-mail: firstname.lastname@example.org, email@example.com