AbstractManeuvering an industrial robot to avoid a collision with obstacles in real time involves not only the fast obstacle detection and descripton, but also fast decision making. The problem is complicated since noa prioriknowledge about obstacles is assumed. In addition, they may appear in the robot's path unexpectedly. To detect and describe the three‐dimensional obstacles, stereo cameras are used to collect environmental images. Through the top view of the workspace, the cameras furnish the silhouette as well as heights of obstacles. To speed up the image processing, the pixel array is grouped into patches and the maximum height of each patch is determined. TO simplify the obstacle description in the computer, a “pillar” model of the bounding polyhedra is constructed. Fast decision making is accomplished by structuring a finite number of possible collision avoidance paths. Path feasibility is determined at the “module aisle” level while optimization is performed at the subpath level so that the magnitude of processing effort is reduced from the order of 6