Within the autonomous robotics literature, bio-inspired models of navigation in organisms (e.g. rats) usually rely on instrumental conditioning processes based on the learning of associations between places in the environment and navigation actions leading to rewarded goal places. This paper presents a neural-network model capable of solving navigation tasks on the basis of Pavlovian conditioning processes (`autoshaping') which allow transferring innate approaching behaviours from biologically salient stimuli (e.g., food) to neutral stimuli (e.g., a landmark seen from far away and close to the food). The overall architecture and functioning of the model is biologically constrained on the basis of relevant neuroscienti¯c anatomical and physiological knowledge on amygdala, nucleus accumbens, and ventral tegmental area. The model is tested with a simulated robotic rat engaged in autoshaping and devaluation experiments. The results show that, although the model allows solving only simple navigation tasks, it produces fast learning and a °exible sensitivity of behaviour to internal states typical of Pavlovian processes. The model is also important for the investigation of adaptive behaviour in general as it clari¯es the nature of some Pavlovian core mechanisms which play a key role in several forms of learning.
