This paper presents an approach for learning object affordances from a human demonstrator. One of the major problems for a robotic agent acquiring a deeper understanding of object affordances is grounding the semantic in its own sensory input. Being able to understand what in the representation of an object makes the object afford the action opens up for more efficient manipulation, interchange of objects that visually might not be similar, transfer learning, as well as removing the need for object categorization. Our approach uses a metric learning algorithm that learns a transform that encourages objects that affords the same thing to be close in space. We regularize the learning such that irrelevant features are penalized, allowing the agent to link what in the sensory input caused the object to afford the action. From this we show how the agent can explain what in the object is specific for an affordance, how the transform can be used to compare how similar different affordances are and how an agent can interchange objects when the object learned on is not available.
\begin{comment}
- state the problem
- say why it is interesting
- say what your solution achieves
- say what follows from your solution. \end{comment}