We present a real-time crowd simulation approach based on reinforcement learning (RL), addressing congestion prevention in confined spaces. We learn a local navigation policy that uses compact, fast-to-compute per-agent observations of a small set of neighbors, including their desired directions. Alongside goal progress and inter-agent spacing, we reward agents for waiting when neighbors ahead pursue similar goals. This formulation fosters global self-organization from purely local interactions. Preliminary results show reduced congestion and consistent goal attainment for large crowds with hundreds of agents.