In recent times the presence of vision and robotic systems in industry has become common place, but in spite of many achievements a large range of industrial tasks still remain unsolved due to the lack of flexibility of the vision systems when dealing with highly adaptive manufacturing environments. An important task found across a broad range of modern flexible manufacturing environments is the need to present parts to automated machinery from a supply bin. In order to carry out grasping and manipulation operations safely and efficiently we need to know the identity, location and spatial orientation of the objects that lie in an unstructured heap in a bin. Historically, the bin picking problem was tackled using mechanical vibratory feeders where the vision feedback was unavailable. This solution has certain problems with parts jamming and more important they are highly dedicated. In this regard if a change in the manufacturing process is required, the changeover may include an extensive re-tooling and a total revision of the system control strategy (Kelley et al., 1982). Due to these disadvantages modern bin picking systems perform grasping and manipulation operations using vision feedback (Yoshimi & Allen, 1994). Vision based robotic bin picking has been the subject of research since the introduction of the automated vision controlled processes in industry and a review of existing systems indicates that none of the proposed solutions were able to solve this classic vision problem in its generality. One of the main challenges facing such a bin picking system is its ability to deal with overlapping objects. The object recognition in cluttered scenes is the main objective of these systems and early approaches attempted to perform bin picking operations for similar objects that are jumbled together in an unstructured heap using no knowledge about the pose or geometry of the parts (Birk et al., 1981). While these assumptions may be acceptable for a restricted number of applications, in most practical cases a flexible system must deal with more than one type of object with a wide scale of shapes. A flexible bin picking system has to address three difficult problems: scene interpretation, object recognition and pose estimation. Initial approaches to these tasks were based on modeling parts using the 2D surface representations. Typical 2D representations include invariant shape descriptors (Zisserman et al., 1994), algebraic curves (Tarel & Cooper, 2000), 2 Name of the book (Header position 1,5) conics (Bolles & Horaud, 1986; Forsyth et al., 1991) and appearance based models (Murase & Nayar, 1995; Ohba & Ikeuchi, 1997). These systems are generally better suited to planar object recognition and they are not able to deal with severe viewpoint distortions or objects with complex shapes/textures. Also the spatial orientation cannot be robustly estimated for objects with free-form contours. To address this limitation most bin picking systems attempt to recognize the scene objects and estimate their spatial orientation using the 3D information (Fan et al., 1989; Faugeras & Hebert, 1986). Notable approaches include the use of 3D local descriptors (Ansar & Daniilidis, 2003; Campbell & Flynn, 2001; Kim & Kak, 1991), polyhedra (Rothwell & Stern, 1996), generalized cylinders (Ponce et al., 1989; Zerroug & Nevatia, 1996), super-quadrics (Blane et al., 2000) and visual learning methods (Johnson & Hebert, 1999; Mittrapiyanuruk et al., 2004). The most difficult problem for 3D bin picking systems that are based on a structural description of the objects (local descriptors or 3D primitives) is the complex procedure required to perform the scene to model feature matching. This procedure is usually based on complex graph-searching techniques and is increasingly more difficult when dealing with object occlusions, a situation when the structural description of the scene objects is incomplete. Visual learning methods based on eigenimage analysis have been proposed as an alternative solution to address the object recognition and pose estimation for objects with complex appearances. In this regard, Johnson and Hebert (Johnson & Hebert, 1999) developed an object recognition scheme that is able to identify multiple 3D objects in scenes affected by clutter and occlusion. They proposed an eigenimage analysis approach that is applied to match surface points using the spin image representation. The main attraction of this approach resides in the use of spin images that are local surface descriptors; hence they can be easily identified in real scenes that contain clutter and occlusions. This approach returns accurate results but the pose estimation cannot be inferred, as the spin images are local descriptors and they are not robust to capture the object orientation. In general the pose sampling for visual learning methods is a problem difficult to solve as the numbers of views required to sample the full 6 degree of freedom for object pose is prohibitive. This issue was addressed in the paper by Edwards (Edwards, 1996) when he applied eigenimage analysis to a one-object scene and his approach was able to estimate the pose only in cases where the tilt angle was limited to 30 degrees with respect to the optical axis of the sensor. In this chapter we describe the implementation of a vision sensor for robotic bin picking where we attempt to eliminate the main problem faced by the visual learning methods, namely the pose sampling problem. This paper is organized as follows. Section 2 outlines the overall system. Section 3 describes the implementation of the range sensor while Section 4 details the edge-based segmentation algorithm. Section 5 presents the viewpoint correction algorithm that is applied to align the detected object surfaces perpendicular on the optical axis of the sensor. Section 6 describes the object recognition algorithm. This is followed in Section 7 by an outline of the pose estimation algorithm. Section 8 presents a number of experimental results illustrating the benefits of the approach outlined in this chapter.