Our contribution tackles the problem of learning to understand anaphor
ic references in the context of robotic machine learning; e.g. Get the
large screw. Put it in the left hole. Our solution assumes the probab
ilistic theory of learning spelt out in earlier publications. Associat
ions are formed probabilistically between constituents of the verbal c
ommand and constituents of a presupposed internal representation. The
theory is extended, as a first step, to anaphora by learning how to di
stinguish between incorrect surface depth and the correct tree-structu
re depth of the anaphoric references.