When manipulating objects in everyday activities, humans make decisions by grounding them in expected consequences. For example, when considering the problem of choosing a location from which an object should be manipulated, then locations are preferred from which the grasping action will likely succeed. The framework of Action-Related Places that is presented in this thesis applies this paradigm of natural decision making to robotics, by computing the expected success of manipulation actions for different robot positions. This enables the robot to find optimal manipulation places. Action-Related Places are effective because they are tailored towards the robot's skills, which is achieved by developing internal models of successful grasping through experience-based learning. Action-Related Places are flexible and are able to consider and optimize a broad range of manipulation constraints.
«
When manipulating objects in everyday activities, humans make decisions by grounding them in expected consequences. For example, when considering the problem of choosing a location from which an object should be manipulated, then locations are preferred from which the grasping action will likely succeed. The framework of Action-Related Places that is presented in this thesis applies this paradigm of natural decision making to robotics, by computing the expected success of manipulation actions fo...
»