● For detecting the scene (e。g。, a carrier or box filled randomly with workpieces) typically laser-based sensor systems are used for acquiring a sufficiently dense point cloud。 The object pose detection  then is considered as a combinatorial optimization prob- lem for which a construction heuristic is applied。 For this heuristic tree search, a finite set of possible workpiece poses is initially derived from the search space。

● In order to use a decision tree, the elements of   the

search set are split into two components: The first component describes a point of interest (POI) in the  search  space which  is part  of the   workpiece

Fig。54。12a,b Car production (a) usually follows the illustrated steps along the assembly line: Stamping of the metal sheet into plates, fixing and alignment of the plates on trays, spot welding, painting the car body, and final assembly of the car body (doors, dashboard, windscreens, power-train seats, and tires)。 Car factories can host over 1000 industrial robots working typically three shifts per day (courtesy of PSA Peugeot Citroën, Paris and Art Movies, Paris)。 The Audi plant in Ingolstadt Germany (b) is highly automated。 The picture shows spot welding robots along the body-in-white transfer line。 Trays carrying car bodies pass through the robot garden

surface。 The second component describes possi- ble workpiece poses relative to a POI。 The partial search quantities obtained thereby have a significant lower complexity compared to the original search set, since the points of interest can provide a con- straint on relative workpiece poses, thus restricting their assumed freedom of movement。

● Typical  tree  search  strategies  such  as   best-first

search can be used。 In that case best-first search explores the search tree by expanding the most promising nodes first。 These nodes are chosen according to a heuristic evaluation score, repre- senting the  estimated  distance  from  the  node to a solution。

● Final  evaluation  of  the  workpiece  poses  is pro-

vided by a six-dimensional (6-D) Hough voting procedure, i。 e。, a generalized Hough transform。 The features used for Hough voting are sensor measure- ments located relative to a POI。 For all possible constellations of a sensor measurement relative  to a point of interest, a probabilistic statement about possible workpiece poses can be made。 Through the superposition of all probability statements, solu- tion candidates can be formed, which are subjected to a statistical test based on a quality rating。 The obtained quality rating along with a given level of significance is used in order to decide about the ac- ceptance of a workpiece pose。

The method is able to locate three to four work- pieces on average  within  0:5—2 s,  using  a  stan- dard desktop computer。 Moreover, robot grippers are equipped with a seventh axis to allow grasping parts

from corners of the bin。 Furthermore, grippers should be formed in such a way that they may  reach  deep into the bin so that they offer only minimal collision volumes。

54。4。2 Welding

Welding is a manufacturing process that joins materi- als by applying heat, sometimes with pressure。 Usually, workpiece material is melted at the process location of- ten with additional filler material。 Typical robot-based welding processes are spot  welding,  particularly  in car body assembly, and gas-shielded metal arc weld- ing (GMAW)。 With increasing compactness of laser sources and robot motion accuracy, laser welding is in the rise。