Runway Zero
RL-trained LLMs recover airport crises that rule-based systems cannot solve end to end.
Space Redeployed to preserve GPU; switched to FREE TIER
Link to Vercel Site - https://project-2pdc2.vercel.app/Airlines still depend on highly trained operations-control teams when disruption cascades across aircraft, crews, passengers, runways, airline economics, and fairness. Runway Zero turns that unsolved end-to-end recovery problem into an OpenEnv environment where LLM agents learn from verifiable rewards.
Hugging Face Space
The Space runs the OpenEnv environment; the large-model training evidence is linked here.
The public Space keeps the reset, step, state, and close API live for evaluation. The 7B, 14B, 31B, and 120B LLM training runs were executed as GPU GRPO jobs, then exported as adapters, trainer states, replay traces, and plots. Keeping those models loaded inside a public demo Space would make the judge experience slow and unreliable, so this page opens the exact replay UI and links the notebooks and artifacts used to verify training.
Hackathon Story
Most benchmarks reward planning. Runway Zero rewards recovery.
Static schedules are easy. The real skill is what happens when fog hits Delhi, Mumbai loses a runway, Bengaluru slots become political, crew legality collapses, passengers are stranded, and every airline asks Tower Central to favor them. Existing tools support pieces of this work; the complete end-to-end recovery problem remains human-led.
Operations Recovery
Fog, runway debris, and aircraft faults test whether the model can make safe dispatch decisions.
Passenger-Aware Recovery
Connections, stranded passengers, emergency arrivals, and gate failures turn delay into human cost.
Economic Multi-Agent Control
IndiGo, Air India, Akasa Air, and SpiceJet negotiate slots while Tower Central preserves fairness.
IndiGo Crisis Replay
A December 2025-style crew availability crisis shows how RL recovery could reduce mass cancellations.