We propose a path optimization method for manipulation. The method tries to minimize path length in the configuration space of the whole system (robots and objects). The problem is cast into a quadratic program with linear constraints. The current path is shorten iteratively. When a collision is detected along an iteration, the method backtracks to the previous collision-free path, adds a linear constraint that prevents the bodies in collision to come close to each other and starts again. The algorithm stops when the global minimum under the collision constraints is reached. The method is run on three different benchmarks, using the real geometric models of the robots Baxter, UR-3 and PR-2.