Crowd simulation becomes one of the active research over the last three decade. Crowd simulation research is part of research about human behavior. One of research focus on crowd simulation was simulation at disaster evacuation. People who trapped in disaster usually has different behavior and abilities when evacuating. So the simulation looks more real, there are several simulations scenario used when disaster simulation like invariant speed and relation between virtual agents scenario. In this research will be used agent-based reciprocal velocity obstacle method. The agent-based method used because applying a microscopic approach. With agent-based method then each virtual agents can have different behavior and abilities. Optimal reciprocal collision avoidance(ORCA) which is the best version of Reciprocal velocity obstacle(RVO) used as virtual agents collision avoidance in crowds simulation. Experiment result demonstrates the simulation of hundreds of agents in a three evacuation scenario. Leader-follower relations on heterogeneous agents result in the possibility of collisions between agents getting bigger with the number of collisions is 7781 collisions. The longest simulation time is 151 seconds with the smallest FPS number is 33 FPS.