Real manufacturing environments are cluttered spaces, where the environment matches its digital model (the CAD model) only to a certain degree. Robot systems, which autonomously execute tasks in these environments should combine CAD and vision data in order to successfully carry out their tasks. This way the model data (coming from the CAD model) can be corrected by the sensor data (coming from computer vision) to properly reflect the real environment. In this paper, a novel method for autonomous task-space exploration is presented, which is based on space exploration and object recognition. During the exploration of the environment, the CAD data is used to determine the initial belief of the newly gained information. After constructing the environment model, a task planner based on a geometric contact analysis and symbolic inference computes the necessary manipulation sequence. To compute the disassembly space, a novel time-efficient algorithm is proposed. The theoretical aspects are applied to disassembly sequences. A variety of heuristics for computing a local optimal sequence of the required manipulation steps and the corresponding paths is also presented in the paper. To plan the paths well-known global path planners are applied with a step-size control which reduce the planning effort. The approaches are experimentally validated and the results are discussed.