The authors describe techniques that generate multiple interpretations from dense range images of piles of unknown objects and methods that use physical law, such as object stability, to rank the interpretations. Each of the interpretations completely accounts for the observed range data, but the interpretations differ in the ways visible portions of objects are extended into the occluded portions of the scene. Experiments with 100 range images indicate that the techniques are fairly robust when the scenes consist of shapes that are approximately prismatic or cylindrical. These techniques are based on novel approaches in several key areas, including explicit use of sensor geometry, generic shape models to synthesize scene descriptions, spatial-reasoning techniques that incorporate knowledge about the laws of physics, direct estimation of the physical properties of the objects in the scene, and detection and refinement of descriptions of approximately planar or cylindrical surfaces.>

The notion of uncertainty manipulation in robot navigation is applied to the task of model-based object recognition. Based on the fact that every detected object feature in an image must be associated with some uncertainty, the authors show how to iteratively refine the estimate of pose uncertainty of a hypothesized object when a new feature is matched, and how to dynamically determine a proper search window to look for the next model feature. The advantage of this method over the conventional static search window is demonstrated in an actual object recognition task.>

Three utility measures that can be used to judge the effectiveness of features in a model-based matching strategy are defined: detectability, reliability, and error rate. With the probability framework, it is shown how it is possible to analytically estimate the reliability of a hypothesis using the measures of the individual features. Based on these measures, the authors have formulated and experimentally verified the matching cost (expressed as execution time) of locating an instance of a model in an image. An algorithm is presented for the selection of optimal seed features of the model using the estimated cost of criterion.>

An experimental robotic ply lay-up workcell whose goal is to automate the lay-up of a broader class of small composite parts is described. Specialized robotic end effectors and new sensing and control techniques are discussed. The feasibility of laying up plies of prepreg composite material onto a simple mold shape is demonstrated.>

Sensing and reasoning techniques for the data-driven description of objects in unstructured environments are given. One of the techniques presented is for extrapolating the shape of individual objects using symmetries, while the other is applicable to objects jumbled in a pile. No a priori object models are needed for these techniques. A hypothesize and test approach is used to infer complete shape information from range images that represent partial knowledge about an object's shape. Using explicit knowledge about the geometrical relationship of the sensor to the scene, hypotheses are verified to ensure that they do not violate physical laws such as object transparency and solidity. In addition, basic physical rules such as stability and contact are used to improve the estimates of the shapes of objects that may be partially visible or partially outside the field of view of the sensor.>