A great deal of A.I. research has focused on teaching computers to identify lists of objects and people. The Multi-sensory Autonomous Cognitive Systems (MACS) project uses a different paradigm - affordance learning. Instead of identifying a specific object as a hammer, an affordance-based system learns the parameters of what makes a hammer useful for hammering. It needs a shaft for leverage, a weight at the end and a flat surface for hammering. Then, if the robot needs to find something with which to hammer, it wouldn't be limited by a narrow visual recognition algorithm for a hammer. It could search for any object suitable for the purpose.

The only given parameters in the Kurt3D test stated that a door switch could be activated by placing a certain weight on a pressure sensitive plate. Kurt3D was able to examine the room, identify an appropriate object, pick it up, place it on the plate, and move through the open door. Photo by: Fraunhofer AIS.

What Can I, Robot, Do With That? [Science Daily]