This paper presents a novel, decentralized, control-theoretic approach to address collision avoidance for multi-robot systems. We create a virtual obstacle at the mean position of the robots. A control is be designed such that each robot will avoid the closest obstacle when a collision is possible. The closest obstacle can be the virtual obstacle or the nearest robot. We present two such control laws. The first assumes perfect knowledge of the velocities of all nearby robots and can allow a saturated velocity input for each robot. In practice, the velocities of the other robots are hard to measure or estimate precisely. Therefore, the second control law removes the assumption of known velocities based on a high-gain, robust control scheme. We prove the first control scheme is globally asymptotically stable, and the robust control law is globally uniformly ultimately bounded. To verify the effectiveness of the proposed approach, Monte Carlo simulations and experiments have been conducted.