Bayesian learning of tool affordances based on generalization of functional feature to estimate effects of unseen tools

To address the problem of estimating the effects of unknown tools, we propose a novel concept of tool representation based on the functional features of the tool. We argue that functional features remain distinctive and invariant across different tools used for performing similar tasks. Such a representation can be used to estimate the effects of unknown tools that share similar functional features. To learn the usages of tools to physically alter the environment, a robot should be able to reason about its capability to act, the representation of available tools, and effect of manipulating tools. To enable a robot to perform such reasoning, we present a novel approach, called Tool Affordances, to learn bi-directional causal relationships between actions, functional features and the effects of tools. A Bayesian network is used to model tool affordances because of its capability to model probabilistic dependencies between data. To evaluate the learnt tool affordances, we conducted an inference test in which a robot inferred suitable functional features to realize certain effects (including novel effects) from the given action. The results show that the generalization of functional features enables the robot to estimate the effects of unknown tools that have similar functional features. We validate the accuracy of estimation by error analysis.