Adjustably Autonomous Robotic Agents

Building autonomous robotic agents that interact with the real-world is a complex and difficult task and several distinct paradigms are used to develop such machines. In the past decade there has been an exploration of approaches inspired by evolutionary and/or neurodynamical principles. [1, 2, 3, 4]. In earlier work, we have examined architectures which support agent-based design of such robotic systems [5]. But, whilst there have been successes in developing systems that exhibit low levels of performance e.g. wall following, safe exploration, map building, it is proving more difficult to develop systems that succeed in high level tasks which depend upon machines being capable of exhibiting (many) different behaviours. One of the motivations for building autonomous robots in which the control system is designed using neurological principles is that such systems are fault tolerant, their performance degrading gradually. However, we have found that one problem with such systems is the initiation of a behaviour in an ’unintended ’ fashion. In the psychological literature, these events are called ’capture errors’; they may arise through incomplete training of neurally-based control system(s) or through ’damage ’ to the control system so that a learnt association is altered [6] (c.f. a stroke in humans)