Instructing production equipment on a shop floor in mid 1990’s is still done using concepts from the 1950’s, although the electro-mechanical devices forming logical gates and memory cells are nowadays replaced by computers. The gates and memory cells are now presented in graphical editors, and there is support for defining new building blocks (gates) handling complex data as well as for advanced graphical monitoring of a production process. The progress during the last 40 years concerns the implementation of the control system, not the instruction of the control system. Although this progress has increased the flexibility of the equipment, it has not provided any support for error recovery. The contribution of this thesis is twofold. Firstly, it presents Aramis - A Robot and Manufacturing Instruction System - which extends the traditional instruction taxonomy with notions developed in computer science. Aramis uses a graphical task specification language which operates on an abstract model of the plant, a world model. The essence of the differences between Aramis and current practice is the usage of abstraction layers to focus on different aspects of the instruction task, and that Aramis retains much specification knowledge about the controlled plant, knowledge which can be reused for other purposes such as error recovery. Secondly, the problems of error recovery is investigated, and a proposal for how to provide recovery support in a system structured such as Aramis is presented. The proposal includes the generation of a multitude of possible restart points in a task program, and it uses a planning approach to support the modification of the current state of the machinery to the ”closest” restart point.