As children start attending school they are more likely to face situations where they have to autonomously learn about novel object categories (e.g. by reading a picture book with descriptions of novel animals). Such autonomous observational category learning (OCL) gradually complements interactive feedback‐based category learning (FBCL), where a child hypothesizes about the nature of a novel object, acts based on his prediction, and then receives feedback indicating the correctness of his prediction. Here we tested OCL and FBCL skills of elementary school children and adults. In both conditions, participants performed complex rule‐based categorization tasks that required associating novel objects with novel category‐labels. We expected children to perform better in FBCL tasks than in OCL tasks, whereas adults to be skilled in both tasks. As hypothesized, in early‐phase learning children performed better in FBCL tasks than in OCL tasks. Unexpectedly, adults performed somewhat better in OCL tasks. Early‐phase FBCL performance in the two age groups was matched, but the OCL performance of adults was higher than that of children. In late‐phase learning there was only an age group main effect (adults > children). Moreover, performance in post‐learning categorization tasks, that did not require label recollection, indicated that in FBCL tasks children were likely to directly learn the associations between an object and a category label, whereas in the OCL tasks they were likely to first learn which feature‐dimensions were relevant. These findings shed light on developmental changes in cognitive control and learning mechanisms. Implications for educational settings are discussed. We tested autonomous observational category learning (OCL) and feedback‐based category learning (FBCL) skills of elementary school children and adults. Early‐phase FBCL performances in the two age groups were matched, but the OCL performances of adults were higher than those of children. Performances in post learning categorization tasks indicated that in FBCL tasks children directly learn the associations between an object and a category label, whereas in the OCL tasks they first learn which feature‐dimensions were relevant.