This paper describes a revised version of ROBOPLAN, a goal-oriented robot task planning system for automatic generation, decomposition and execution of high-level robot plans for assembly. It emphasizes its new features, i.e., modularity, formal definition of the task, robust plan synthesis, and execution of each assembly step. A task definition language allows a formal description of the robot universe and the assembly task to be input to ROBOPLAN. The expert task planner is a non-linear backward chaining problem solver, using a goal driven depth-first strategy. The implemented search strategy has been tested in the assembly domain, but it could be used in other domains where planning is needed. The motion planner provides a non-optimal, safe robot trajectory; collision free path planning has not been included yet. A robot executable code is generated for each assembly step and monitored in real time. The error detection and recovery capability of the system is rather limited at present, since no sensors are used. The initial implementation of the system has been tested and evaluated on the assembly of a DC motor. The potential of extending this planning framework to other applications is also discussed.