Learning of usage of tools based on interaction between humans and robots

To perform a tool-using task by manipulating some target object, a robot needs to a) determine the desired effects of the task b) determine a suitable tool c) generate the correct position and orientation in which tool should be placed relative to the target object and d) determine the action to subsequently manipulate the tool after it is placed. Thus, learning to use tool requires robot to determine casual dependencies that exist among the spatial constraints of the environment, desired effects, functional features of the tool, actions and structural constraints of the tool. But learning of these casual dependencies by performing tool manipulation tasks using self-exploration or pre-programming by user is computationally expensive and resource intensive process. Thus, to address such a problem, learning of tool use in an online, incremental and interactive manner is proposed, where robots learn tool-use using trial-and-error-and-interaction loop. To deal with uncertainties in the domain knowledge, robot learning and inference process, probabilistic semantics of Bayesian network(BN) is used. BN allows making probabilistic queries to human user based on internal state of the robot and also allows incorporating user feedback with a degree of confidence. The concept of interaction based learning and experimental scenarios are presented.