Abstract : Virtual reality environments are becoming increasingly large and complex and real-time interaction level is becoming difficult to stably insure. Indeed, because of their complexity, detailed geometry and specific physical properties, these large scale environments create a critical computational bottleneck on physical algorithms. Our work focused on the first step of the physical process: the collision detection. These algorithms can sometimes have a quadratic complexity. Solving and simplifying the collision detection problem is integral to alleviating this bottleneck. Hardware architectures have undergone extensive changes in the last few years that have opened new ways to relieve this computational bottleneck. Multiple processor cores offer the ability to execute algorithms in parallel on one single processor. At the same time, graphics cards have gone from being a simple graphical display device to a supercomputer. These supercomputers now enjoy attention from a specialized community dealing solely with physical simulation. To perform large scale simulations and remain generic on the runtime architecture, we proposed unified and adaptive mapping models between collision detection algorithms and the runtime architecture using multi-core and multi-GPU architectures. We have developed innovative and effective solutions to significantly reduce the computation time in large scale environments while ensuring the stability and reproducibility of results. Our models cover the global collision detection pipeline, focusing both high and low level algorithms. Our models based on multi-core, GPU and multi-GPU combine different techniques of spatial subdivision algorithms based on topology and load balancing techniques based on data stealing. Our hybrid solution enables us to increase space and computing time within large scale virtual environments. The coupling of these new algorithms led us to develop two models of dynamic algorithmic adaptation based (or not) on off-line precomputed scenarios. Finally, it became necessary to add a new dimension to the collision detection pipeline taking into account of the architecture for optimal execution. With this formalism, we proposed a new pipeline of collision detection with a granularity of parallelism on multicore processors or multi-GPU platforms. It enables simultaneous execution of different stages of the pipeline and a parallel internal to each of these steps.