Robot learning operates at the crossroads of disciplines such as machine learning, robotics engineering, and developmental robotics for lifelong learning. Robot skills can be divided into four categories: sensorimotor (locomotion, grasping); interactive (joint manipulation of an object); linguistic; and autonomous self-exploration or exploration through guidance from a human teacher.

Therefore, robot learning can be closely related to subject areas such as adaptive control, for improving sensorimotor skills via dynamically adapting controllers; reinforcement learning, for understanding, taking actions, and planning; and developmental robotics, for more degrees of autonomous learning modalities such as those existent in human children, where lifelong learning is expected to be cumulative and of progressively increasing complexity.

In this context, the paper addresses the areas of computer vision and natural language understanding, with emphasis on robot learning, exceptionally well. In particular, it provides an excellent starting point for someone to do research in this area. Also, from a teaching point of view, it provides an excellent reading list for postgraduate students.

Although the paper is well written, it takes a long time to reach the point when it concentrates on computer vision and natural language understanding specifically for robots. Instead of concentrating on related work about the integration of computer vision and natural language for robots, the authors first take two separate long journeys into the semantics of natural language processing (NLP) and computer vision and/or image annotation. There is a huge body of related work about the semantics in these two separate areas, and there are many other survey papers. Therefore, the reader may feel a bit disappointed having gone through this survey, particularly if he or she is already familiar with aspects of semantic computing in NLP and computer vision.