It is said that when Magritte was asked why the image he had painted was
not a briar pipe, he responded "Because I can't use it as a pipe".
We believe that in order to represent and recognize objects, it is important to address not only some of the intrinsic properties of the object but also, how it is used: its functionality.
The observability of the interaction is crucial not only for controlling the active and interactive observers but also for monitoring the functional investigation. By considering the mechanics of an interaction and its results we can investigate the functional features that lead to a successful interaction.
Our current investigation focuses on manipulatory interactions and thus far the functionality we have addressed is that of piercing. In the experiments we vary the material properties of tool and target objects (e.g. shape and hardness). Results from the force, position and vision observers, which allowed us to identify tools having the functionality of piercing are combined in the form of force and shape maps. These maps proved a coherent frame-work for combining experiments carried out employing different tools in the same context. Further analysis incorporating other intrinsic features will enable us to better recognize the functional features which distinguish the objects and define criteria of task performance. The current implementation of the system employs two Puma-560 robots. One is equipped with a Lord force-torque sensor and handles the tool. The other holds a camera to track and monitor the progress of the task as the tool comes into contact with the target object. We envision that such a system would be extended to have at its disposal a score of investigative procedures and hence be able to classify and recognize an object based on its functional attributes. Thus, the active investigation of the functional features of an object provides a decisive step towards its functional recognition and its role in the scene.
The integration of other sensor modalities to adapt to a changing environment and for recognizing the importance of other features is currently being investigated. The type of observation we are planning to incorporate is that of bending. A new observer will be incorporated as part of the control if its observation provides the ability to better qualify the interaction and to identify instances which, upon verification, were incorrectly classified. In addition we wish also to focus on possible deformations that the target object will be subjected to. When considering cutting, for instance, it is necessary to observe how the target object is being affected by the tool in order to identify a successful interaction.
The system provides output in two forms. A running time output is provided to the user in the form of graphs indicating the state of the execution of the task as perceived by the various observers. Each execution provides force profiles which are then combined with information about the shape to construct force-shape maps. These, in turn, allow a classification of the tool given the recovered structural properties of the object and the dynamics of the interaction. Thus, these maps provide the second type of output.
Future extensions of our investigation will focus on carrying out other
functional interactions as well as further developing the concept of the
extracted force-shape maps.
References and on-line Publications.
Tools' functionality in the Stone Age.
The world of Function,
Recognition.