Velocity-based models belong to the category of microscopic crowd simulation models. They recently appeared in the crowd simulation literature. They mathematically formulate microscopic interactions as a function of agents' states and their derivatives. In the case of collision avoidance, this property provides agents with the ability to produce anticipated smooth reactions, with great impact on simulation results. This paper is made of 5 sections. We first describe the main principles of velocity-based models. We then describe an experiment that tends to prove that velocity-based models are founded, at least in the basic case of collision avoidance between two walkers. Next, we describe 3 models, the Paris model, the Tangent model and the Vision model that were successively proposed by our INRIA group for crowd simulation. We emphasize their differences to allow comparison. We finally discuss the benefit of velocity-based approaches, and what are the problems raised by these approaches and to be addressed in future work