Affordance

An affordance is what an object or environment lets you do. A handle affords grasping, a button affords pushing, a flat surface affords placing something on it. The concept, borrowed from psychology, is useful in robotics because it frames perception in terms of possible actions rather than just object identity.

Predicting affordances tells a robot not only what something is, but how it can interact with it.

Affordance prediction models take perception input, such as an image or point cloud, and output where and how interaction is possible, for example graspable regions on an object or a surface that can support placement. These predictions then guide grasping, manipulation, and planning. Affordances can be learned from demonstrations of interaction or from labeled interaction data.

Recognizing an object is not enough for a robot that has to act on it, and affordances supply the missing link between perception and action. They are central to manipulation and to general-purpose robots that must interact with unfamiliar objects in sensible ways.

Object recognition identifies what something is, while an affordance describes what actions it makes possible, which is what a robot needs to interact.

They connect perception directly to action, guiding where and how a robot can grasp, push, or otherwise interact with objects.

Grasping, Embodiment, Policy, End-effector.