Autonomous robots motivated to learn
In an unknown and continuously changing environment, an autonomous robot must be able to react instantaneously to changes and unexpected events in order to avoid collisions and update its maps. For successful navigation the robot is necessary to respond based primarily on its immediate sensory information and secondarily on internal maps of the surrounding environment. Such intelligent behaviour is, besides other prerequisites, a direct consequence of the system's architecture combined with a learning and adaptation scheme, individually shaped with respect to the experience gained. Under the coordination of the University of Bonn, the SIGNAL project partners sought to design a systemic architecture, enabling autonomous robots to grow up through a sequence of learned capabilities. The systemic architecture proposed was organised in a hierarchy of layers, which correspond to distinct levels of the robot's capabilities and are divided according to the information pathway. Sensory input from primary sensors is processed to derive information about the surrounding environment through a cascade of pattern recognition and categorisation modules, until symbolic information is generated. The flow of information is directed through multiple layers to provide for basic sensory data to more complicated information processing tasks, before results are made available to actuators modules. Adequate commands for the robots' actuators are then generated with respect to the task that the intelligent system has to perform as well as to the "internal state" of the system. While the robot is moving through its surrounding environment, interconnection graphs of successive sensory items representing objects of the real world are being built using the successive actions performed. These topological associations of the surrounding environment can provide the internal maps needed for navigation and action planning. Moreover, individual modules within each system layer are adaptive and capable of learning their respective functionality through paradigms originating from neural networks, machine learning, artificial immune systems and psychology learning. In contrast to conventional systems that are designed to acquire capabilities in one bunch, the sensory-actuatory mapping method allows technical artefacts to acquire sub-functionalities through learning.
