Studies of searching behaviour in bees by Cartwright and Collett led to a computational model of short-range insect guidance that has been successfully implemented on real robots. Still, is reliability depends crucially on arriving at a good match between landmarks observed in the goal place and those in any nearby place within the environment. This paper describes an application of this model in a standard office environment, with unprepared landmarks that may occasionally become invisible or that are easily confused. The corresponding approach calls upon a visual chip that perceives colour and the whole height of the visual field, and upon a matching algorithm that uses colour and proceeds globally, using dynamic programming. Together, they lower the risk of spurious landmark matchings and enhance the performance of the algorithm significantly, allowing it to work without a full 360° panorama and to cope with object disappearance. The performance with respect to the original model of Cartwright and Collett is assessed, both in simulation and in experiments with a real robot. Improvements over previous robotic applications of this model, or its variants, are emphasized. Directions for future improvements are indicated.