Runway Zero: Training Agents To Recover After The Plan Breaks
Most agent benchmarks ask whether an LLM can make a plan. Runway Zero asks a
harder question: can it recover when the plan starts failing in real time?
Runway Zero is an OpenEnv-style environment for cascading Indian airport
operations. The agent begins with a normal national flight schedule, then has to
respond as disruptions compound: fog hits Delhi, Mumbai loses runway capacity,
Bengaluru gates jam, an IndiGo aircraft needs maintenance, crews approach
duty-time limits, Hyderabad receives an emergency arrival, and airlines compete
for scarce evening slots.
The agent does not answer with prose. It emits structured JSON actions:
departures, holds, cancellations, aircraft swaps, maintenance requests,
reroutes, passenger compensation, connection protection, and slot negotiation.
Those actions are executed inside the simulator. The environment returns a
decomposed reward over delay, safety, passenger satisfaction, airline money,
fairness, and action validity.
The environment has three levels:
- Operations Recovery: a compact four-airport network focused on safe
departures, arrivals, aircraft readiness, and delay reduction.
- Passenger-Aware Recovery: a larger network with connections, stranded
passengers, emergencies, and satisfaction penalties.
- Economic Multi-Agent Control: ten Indian airports where IndiGo, Air India,
Akasa Air, and SpiceJet compete for slots while Tower Central must stay
neutral.
For the demo, the website replays real simulator traces. It shows a custom
animated India operations board, airport zoom views with runways and gates,
active disruption context, agent negotiation messages, reward bars, speed
controls, and base-model-vs-RL-trained comparison rows.
Training evidence is included at three levels. First, deterministic baselines
and a local all-stage RL controller generate plots and replay traces. Second, a
local TRL/GRPO smoke run proves the LLM reward loop executes against the real
environment. Third, hosted Hugging Face GPU jobs trained four large model
policies with TRL GRPO across all three stages: Qwen2.5-Coder-7B-Instruct,
Qwen3-14B, GPT-OSS-120B, and Gemma-4-31B-IT. Each run uploaded its adapter
artifact bundle and JSON summary to the Hugging Face artifact repo.
Why this matters: deployed agents will not operate in clean, single-turn
prompt-response tasks. They will operate inside systems where resources
disappear, stakeholders disagree, APIs fail, and early decisions change future
state. Runway Zero turns that recovery skill into a measurable training
environment.
