In this paper, we propose a collision avoidance algorithm for multi-vehicle systems, which is a common problem in many areas, including navigation and robotics. In dynamic environments, vehicles may become involved in potential collisions with each other, particularly when the vehicle density is high and the direction of travel is unrestricted. Cooperatively planning vehicle movement can effectively reduce and fairly distribute the detour inconvenience before subsequently returning vehicles to their intended paths. We present a novel method of cooperative path planning for multi-vehicle systems based on reinforcement learning to address this problem as a decision process. A dynamic system is described as a multi-dimensional space formed by vectors as states to represent all participating vehicles' position and orientation, whilst considering the kinematic constraints of the vehicles. Actions are defined for the system to transit from one state to another. In order to select appropriate actions whilst satisfying the constraints of path smoothness, constant speed and complying with a minimum distance between vehicles, an approximate value function is iteratively developed to indicate the desirability of every state-action pair from the continuous state space and action space. The proposed scheme comprises two phases. The convergence of the value function takes place in the former learning phase, and it is then used as a path planning guideline in the subsequent action phase. This paper summarizes the concept and methodologies used to implement this online cooperative collision avoidance algorithm and presents results and analysis regarding how this cooperative scheme improves upon two baseline schemes where vehicles make movement decisions independently.