Autonomous Detection and Experimental Validation of Affordances

We propose a computational formalization of affordances, which is able to consistently combine affordance-related evidence resulting from different observations. We represent affordances as Dempster-Shafer belief functions defined over the space of end-effector poses, which can be combined using uncertain logic in order to allow their hierarchical organization. The primary source of affordance-related evidence is visual affordance detection, which first simplifies the perceived environment into geometric primitives and then evaluates a hierarchical set of affordance definitions based on the available visual information. The resulting belief functions are used as initial affordance hypotheses, which are subject to further investigation and validation. As pure visual affordance detection can fail to properly estimate important preconditions, e.g., the stability of environmental structures, validation experiments are conducted in order to incrementally improve the system belief and the reliability of detected affordances. The proposed formalism is implemented and evaluated in the context of loco-manipulation affordances for humanoid robots using the simulated robots ARMAR-III and ARMAR-4.