Blog.md

Ambulance Green Corridor — OpenEnv Hackathon 2026

Theme: #3.1 — World Modeling / Professional Tasks

One-line summary: We train an LLM to act as an emergency dispatcher + city traffic signal manager, navigating a partially observable city with real-world constraints (gridlock, potholes, accidents, hospital capacity) to get ambulances to the right hospital as fast as possible.

Demo video: Watch on YouTube

---

The Capability Gap We Are Targeting

Current GPS-based emergency preemption systems (like Opticom, used globally) work like this:

Ambulance within 300m of an intersection → turn it green.

That is reactive. It has no awareness of:

• What road quality lies ahead (potholed roads slow the ambulance even on green)
• Whether the nearest hospital is the right hospital for this patient's condition
• Whether heavy traffic on the planned route makes a longer detour actually faster
• Dynamic events mid-journey: accidents, road closures, traffic spikes

The question we ask: Can an LLM reason about the full journey — hospital selection, road quality, live traffic state, and mid-episode events — to get ambulances to the right place faster than any rule?

This is a genuine professional task that requires persistent world modeling across many steps. A rule cannot solve it. A shortest-path algorithm cannot solve it. This environment tests whether an LLM can.

---

Theme Alignment: Why This Is Theme 3.1

Theme 3.1 asks for environments where:

"the model is expected to do real hard work instead of exploiting short-cuts"

Our environment prevents shortcuts in three ways:

1. Toggling already-green signals costs reward. The agent must read signal state before acting — it cannot blindly clear everything.
2. Traffic volume slows the ambulance even on green. The agent cannot just clear signals and assume it will go fast — it must reason about the traffic volume on each segment.
3. The nearest hospital is not always correct. A cardiac patient sent to a trauma centre loses the +300 specialist bonus. But even the right specialist hospital may not be the best choice — if the route to it is gridlocked, clearing signals only gets you 20% speed through dense traffic. A farther hospital with lighter traffic and a lower ETA is the smarter pick. The agent must weigh specialization + distance + live traffic volume simultaneously, and be willing to switch hospitals mid-journey if conditions change.

The agent must maintain a coherent world model across 15–30 steps, update it when dynamic events fire, and make non-obvious decisions that only pay off several steps later.

---

Environment Design

What the Agent Sees (Observation)

Every step, the agent receives a structured observation:

=== EMERGENCY DISPATCH ===
Patient  : (6, 3) | condition: cardiac
Ambulance: (6, 4) | time: 40s / 300s

⚠ DYNAMIC EVENTS:
  [ACCIDENT] at (4,3) — road blocked (severity=0.8)

CURRENT ROUTE → hosp_a (City General)
  ETA=251s | segments=8 | damaged=2 | heavy_traffic=1
  (6,4)→(5,4) | residential | quality=moderate | traffic=45%
  (5,4)→(4,4) | damaged     | quality=POTHOLED | [BLOCKED]

ALTERNATIVES:
  hosp_c (Cardiac Centre) ← specialist match | ETA=130s | damaged=0

SIGNALS — only change WRONG ones:
  (5,4): ns_green | ambulance going north | OK
  (4,4): ew_green | ambulance going north | WRONG — needs ns_green

ACTION FORMAT:
{"hospital_id": "hosp_c", "signal_controls": [{"row": 4, "col": 4, "phase": "ns_green"}], "preferred_direction": null}

What the Agent Does (Action Space)

{
  "hospital_id": "hosp_c",
  "signal_controls": [{"row": 4, "col": 4, "phase": "ns_green"}],
  "preferred_direction": "north"
}

• hospital_id: choose or switch destination at any step (not just at the start)
• signal_controls: override up to 3 signals in the lookahead window
• preferred_direction: hint the routing engine to take a specific turn

Reward Function — Designed to Be Hard to Game

Component	Value	Purpose
Arrival	+1000	Primary objective
Time bonus	+500 max	Rewards speed
Specialist match	+300	Rewards reading patient condition
Red light stop	−20 each	Penalises poor signal management
Unnecessary toggle	−2/−5 each	Core anti-shortcut mechanism
Damaged road traversed	−10 each	Rewards road quality awareness
Successful re-route	+50 each	Rewards dynamic adaptation

The unnecessary toggle penalty is the key design decision. An agent that blindly clears every signal in view scores lower than one that reads the state first. This forces genuine reasoning, not pattern-matching.

Difficulty Levels

Level	Grid	Hospitals	Base Traffic	Events/Step	Time Limit
easy	6×6	2 general	Low (0.1)	5%	200s
medium	8×8	3 mixed	Moderate (0.3)	10%	300s
hard	12×12	5 (1 at capacity)	Heavy (0.5)	15%	400s

---

Training

Model: Qwen/Qwen2.5-0.5B-Instruct + LoRA (r=16, 2.1M trainable params)
Algorithm: GRPO (Group Relative Policy Optimisation) via HuggingFace TRL
Setup: 10 iterations × 4 episodes per iteration, live environment connection

Results

Metric	Baseline (untrained)	After Training	Change
Arrival rate	100%	100%	—
Signal efficiency	11%	100%	+89 pp
Mean reward	1442.6	1445.3	+2.7

What the Numbers Mean

Signal efficiency is the core proof of learning.

The untrained model (11% efficiency) toggles every signal it encounters, including ones already in the correct phase. It treats the action space as "clear everything" — a shallow shortcut.

After GRPO training (100% efficiency), the model learned to:

1. Read current_phase from the observation
2. Compute needed_phase based on the ambulance's direction of travel
3. Only send a SignalControl action when they differ

This is not a trivial pattern. The mapping between ambulance direction and required signal phase differs per intersection and must be reasoned about from the observation text each step.

The training curve shows the characteristic GRPO exploration-convergence pattern:

• Iterations 1: Model arrives (100% arrival) but wastes actions (11% efficiency)
• Iterations 2–4: Exploration phase — arrival drops to 0–25%, model tries aggressive strategies
• Iterations 5–10: Convergence — 100% arrival with 100% signal efficiency simultaneously

---

Live Demo

Demo video: Watch on YouTube
Environment (OpenEnv WebSocket): wss://ajitg25-ambulance-green-corridor.hf.space/ws
Visual simulation: https://huggingface.co/spaces/Ajitg25/ambulance-green-corridor
GitHub (full code + notebook): https://github.com/ajitg25/openEnv-hackathon/tree/main
Training notebook: https://github.com/ajitg25/openEnv-hackathon/blob/main/examples/ambulance_grpo_training.ipynb

Connecting Your Own Agent

from ambulance_env import AmbulanceEnv, AmbulanceAction, SignalControl

async with AmbulanceEnv(base_url="https://ajitg25-ambulance-green-corridor.hf.space") as env:
    # Reset — get patient location, hospitals, initial state
    obs = (await env.reset()).observation

    # Step 1: Dispatch to specialist hospital
    obs = (await env.step(AmbulanceAction(hospital_id="hosp_b"))).observation

    # Step 2+: Clear only wrong-phase signals each step
    while not obs.done:
        controls = [
            SignalControl(
                row=s.row, col=s.col,
                phase="ns_green" if s.ambulance_direction in ("north", "south") else "ew_green"
            )
            for s in obs.lookahead_signals
            if s.phase != ("ns_green" if s.ambulance_direction in ("north", "south") else "ew_green")
        ]
        result = await env.step(AmbulanceAction(signal_controls=controls))
        obs = result.observation

The Three Policies (Shown in Visual Demo)

Policy	Behavior	Signal Efficiency	What It Demonstrates
No control	Does nothing	0%	Pure baseline
Naive	Clears all signals	~11%	Untrained LLM behavior
Smart	Clears only wrong-phase	100%	Trained LLM behavior

---

Why This Matters

Emergency vehicle routing is deployed infrastructure. The gap between "clear one signal 300m ahead" and "reason about the full journey in a partially observable dynamic city" is exactly the gap this environment is designed to close.

Could a researcher write a paper about this? Yes:

"LLM-based adaptive emergency corridor planning under partial observability and dynamic constraints"

That paper does not exist yet. This environment is the training ground for it.

---

Built for the OpenEnv Hackathon India 2026 — Theme 3.1: World Modeling / Professional Tasks