Robotic Affordance Learning: Old Ideas, Recent Developments, and Potential Paths Forward

The notion of affordances as a concept originated in the late 70’s in the field of ecological psychology as founded by Gibson and others [1], and attempts to enable robots to autonomously learn the affordances in their environments, as recent survey papers reveal [2, 3], emerged not long after in the early 90’s, reaching a particularly vibrant zenith in recent years. While many of these works have borne fruit in restricted experimental settings, where the environ- ment and robotic interactions can be controlled and guided towards certain goals, the abiding concern of designing robots capable of general and continu- ous affordance learning remains elusive. Meanwhile, recent experiments using deep neural networks in the fields of machine learning and computer vision [4, 5] have demonstrated that given sufficient data, computational resources, and algorithmic proficiency, impressive results can be achieved when it comes to general representations and inference. This naturally raises many questions regarding the current state of the field of robotic affordance learning and how best to push it forward.