The emergence of action sequences from spatial attention: insight from mammal-like robots. Ben Mitchinson 1 , Martin J Pearson 2 , Anthony G Pipe 2 , and Tony J Prescott 1 1 ATLAS Research Group, The University Of Sheffield, UK 2 Bristol Robotics Laboratory, Bristol, UK. Abstract. Animals display patterns of behaviour that are integrated over space and time. One plausible candidate for the decomposition of control is that behavioural sub-systems compete to command effectors. An alternative hypothesis, inspired by research on primate visual at- tention, is that actions are directed at a sequence of foveation targets selected using a topographic ‘salience map’. In small terrestrial mam- mals, many behaviours are underpinned by foveation, since important effectors (teeth, tongue) are co-located with foveal sensors (microvibris- sae, lips, nose), suggesting a central role for foveal selection in generat- ing integrated behaviour. Here, we investigate control architectures for a biomimetic robot equipped with a rodent-like vibrissal tactile sensing system, explicitly comparing a salience map model for action guidance with an earlier model implementing behaviour selection. Both architec- tures generate life-like action sequences, but in the salience map version higher-level behavioural ‘bouts’ are an emergent consequence of following a shifting focus of attention. Keywords: brain-based robotics, action selection, tactile sensing, be- havioural integration, saliency map 1 Introduction The problem of behavioural integration, or behavioural coherence, is central to the task of building life-like systems [?,?]. Living, behaving systems such as ani- mals display patterns of behaviour that are integrated over space and time such that the organism controls its effector systems in a co-ordinated way and gen- erates sequences of actions that serve to maintain its homeostatic equilibrium, satisfy its drives, or meet its goals. How animals achieve behavioural integration is, in general, an unsolved problem in anything other than some of the simplest invertebrates. What is clear from the perspective of behaviour is that the problem is under-constrained since similar sequences of overt behaviour can be generated by quite different underlying control architectures [?]. This implies that to un- derstand the solution to the integration problem in any given organism is going to require investigation of mechanism in addition to observations of behaviour. In this regard, physical models—such as robots—can prove useful as a means of embodying hypotheses concerning alternative control architectures whose be- havioural consequences can then be measured observationally [1]. Research with