Humans integrate visual and haptic size information in a statistically optimal fashion (Ernst & Banks, 2002). That is, the perceived size is a weighted average of the individual estimates with weights proportional to their inverse variances. More importantly, the fused percept has lower variance then each individual estimate. Fusion of visual and haptic size estimates is reasonable because in the natural environment these cues to an object’s size are highly correlated. The purpose of this study is to investigate whether cue fusion is learned based on the correlation between cues. Therefore, we took naturally uncorrelated cues - the luminance of an object (visual cue) and its stiffness (haptic cue) - and trained 6 subjects for approximately one hour in an environment where these cues were correlated. To test whether training had an effect we compared subject’s discrimination performance before and after training for two intermixed conditions: One condition in which the two cues were consistent with the correlation during training (congruent) and the other condition in which the two cues were anti-correlated relative to the training phase (incongruent). If training had an effect we would predict that the stimuli with congruent cues elicit an improvement in discrimination performance relative to the incongruent condition, because if the cues are fused after training the variance of the combined estimate should get lower. In agreement with our prediction we found a significant interaction between pre- and post-test for the two congruent and incongruent conditions (F[1,5]=20,3; p<0.01). This indicates that subjects indeed picked up the correlation in the training phase and fused the two cues. We conclude that fusion of cues can be learned on a relatively short timeframe based on the statistics of their co-occurrence.