Towards cognitive robotics : robotics, biology and developmental psychology

In cognitive science, modelling has many useful roles, one of which is to inspire attempts to realise an artificially intelligent agent or autonomous robot. However, there still remains a large gulf between the behaviour produced by our best robotic efforts and the richness of behaviour, learning and adaptability so obviously manifest in living systems.

One approach to this problem is to characterise and model cognitive processes whose validity for engineering applications can be assessed through the building of (autonomous) robots working in everyday (non-toy world) environments. This approach will be described together with its emphasis on psychologically tenable theories of learning and cognition.

Starting from a servo-based model of behavioural control proposed by William Powers, this treatment of adaptation and anticipation can be extended and developed for robotics. The basic concept involves negative feedback loops closed by environmental interaction so that the error between an internal reference and a perception is minimised by action. Thus behaviour is the consequence of tracking reference goals by reducing perceptual mismatch. This idea will be discussed and lead on to an extension that offers a framework in which low level reactive behaviour can be integrated with higher-level schemas in a control hierarchy. The relevance for real applications is illustrated by robot grasping experiments where a natural food product is to be grasped by a commercial assembly robot. No prior knowledge of the product shape is available but a grasping algorithm is described that solves this problem by adapting the robot gripper orientation to perceived features of the product without operator intervention.

Building on these ideas, further developments are described involving 3D vision and pattern recognition, 3D registration, fast 3D reconstruction, real time tracking of faces and eyes, and a fully automatic real time 3D face recognition system. The implications of the research are considered for a range of application areas such as 3D CCTV systems, 3D animation and entertainment.

Other branches of this research have explored the scientific basis of robotics experiments, the rules that govern the interaction between a robot and its environment, and how these could be identified and modelled faithfully. One of the first scientific approaches for quantitative modelling of mobile robot behaviour and the evaluation of mobile robot learning and navigation tasks is described.

Finally, the role of developmental psychology in robot learning is discussed, with particular emphasis on the significance of constraints, the growth of competence and identifiable stages in sensory-motor behaviour.