Extraction of multi-modal object representations in a robot vision system