We proposed in [LGA13] an iterative ray-traced collision detection algorithm (IRTCD) that exploits spatial and temporal coherency and proved to be computationally efficient but at the price of some geometrical approximations that allow more interpenetration than needed. In this paper, we present two methods to efficiently control and reduce the interpenetration without noticeable computation overhead. The first method predicts the next potentially colliding vertices. These predictions are used to make our IRTCD algorithm more robust to the above-mentioned approximations, therefore reducing the errors up to 91%. We also present a ray re-projection algorithm that improves the physical response of ray-traced collision detection algorithm. This algorithm also reduces, up to 52%, the interpenetration between objects in a virtual environment. Our last contribution shows that our algorithm, when implemented on multi-GPUs architectures, is far faster.