Interactive Incident Response Control Room — a multi-step agentic RL environment where AI agents monitor surveillance feeds, detect anomalies and escalate incidents in real time.
Agents navigate between 1–4 camera feeds, tracking suspects across locations. Each camera provides different viewpoints of the same incident timeline.
5–15 step episodes with shaped trajectory rewards. Agents must inspect frames, navigate timelines, classify threats, and make escalation decisions.
Rubric-based scoring — no LLM-as-judge. Three difficulty-specific graders with reproducible scores between 0.0 and 1.0.
Sends real CCTV frame images (base64-encoded) to VLM agents. Temporal variation and HUD overlays simulate authentic surveillance footage.