It is well known that path planning for robots with many degrees of freedom is a complex task. That is the reason why the research on this area has been mostly restricted to static environments. This paper presents a new method for on-line path planning for robotic arms in dynamic environments. Most on-line path planning methods are based on local algorithms that end up being inefficient due to their lack of global information (local minima problems). The method presented in this paper avoids local minima by using a two stage framework. The robots react to dynamic environments using a local and reactive planning method restricted to a subset of its configuration space. Since the subset has few degrees of freedom the computational cost of the on-line stage is very low. An off-line stage chooses the subset of the configuration space that minimizes the probability of blockades and inefficient motions.