Note Start here: the Simulation RL Environments article. Explore PortSim and AirOpsEnv, watch the launch videos, and inspect the full AirOpsEnv v3 results: six models across 24 storm dates, recorded rollouts, and downloadable evaluation data.

Note AirOpsEnv v3: three communicating agents run airline operations, crew and customer desks through storm days. Grounded in JFK + Boston flight records, FAA advisories and NWS weather. Built by Edgar Arakelyan (@earakely) with @agutmanstein, a licensed pilot, in collaboration with Scale AI (@scale_ai) and Hugging Face. Built using AgentEnv Framework: https://agentenvframework.com

Note AirOpsEnv v3: 44 training dates, 24 storm evaluation dates and 2 calm controls. Six-model, three-agent leaderboard plus a one-agent comparison: 168 scored v3 runs with recorded actions and messages. Includes tasks, reference plans and replay data. Crew rosters and passenger itineraries are generated from public records.

Note PortSimEnv v2–v5 on AgentEnv, by Edgar Arakelyan (Scale AI): every run replayed on the 3D quay.

Note PortSimEnv v2–v5, by Edgar Arakelyan (Scale AI): the live, marine, wind and lines weeks, references and every recorded run.

Note PortSimEnv v1: the OpenEnv environment, editor, 3D viewer and eval explorer

Note PortSimEnv v1: tasks, the source port calls, eval rollouts

Note PortSimEnv v1 eval: the eval table and every model rollout in 3D