Inspiration
Vietnam already has millions of surveillance cameras, industrial sensors, and camera AI systems, but most of them stop at detection and alerting. When an incident happens — a drowning risk, fire, abnormal movement, infrastructure damage, or a safety violation — the response often still depends on a human operator seeing the alert, understanding the scene, calling the right team, and sending someone to verify it.
That delay inspired us to build AeroGuard Agentic Response: an AI agent system that connects existing cameras and sensors with UAVs, so incidents can be detected, verified on-site, and escalated into a safe first response much faster.
The core idea is simple:
Cameras should not only “see” incidents. They should trigger an intelligent, controlled, and auditable response.
Our target is to reduce the time from signal detection to on-site visual verification and first response to under 120 seconds, while keeping critical decisions under human approval.
What it does
AeroGuard is an agentic UAV coordination system for real-time monitoring and emergency response.
The system receives signals from existing cameras, sensors, or external systems, then uses computer vision and AI agents to:
- Detect potential incidents or abnormal events.
- Understand the scene and estimate urgency.
- Decide whether a UAV should be dispatched.
- Select the right drone, payload, route, and mission plan.
- Verify the event with live aerial footage.
- Trigger safe first-response actions such as visual confirmation, warning, reporting, or human approval for higher-risk actions.
- Generate an incident report with evidence and timeline.
Instead of replacing existing camera infrastructure, the project adds an agentic response layer on top of it.
How we built it
We designed the system as a multi-agent architecture.
At the center is an Orchestrator Agent, which coordinates specialized agents:
- Perception Agent: analyzes camera feeds, object detection, anomaly detection, tracking, and scene understanding.
- Mission Planner Agent: chooses the drone, target location, payload, route, and mission objective.
- Flight Control Agent: sends commands to the UAV system and monitors mission status.
- Critic / Verifier Agent: checks whether the mission decision is reasonable and safe before execution.
- Safety / Guardrail Agent: enforces geofencing, no-fly zones, altitude limits, privacy constraints, and human-in-the-loop rules.
- Reporter Agent: summarizes the mission, evidence, and outcome into a human-readable report.
The system follows an agentic loop:
$$ Signal \rightarrow Detect \rightarrow Reason \rightarrow Plan \rightarrow Verify \rightarrow Act \rightarrow Report $$
For the prototype, we focused on the decision layer: how an AI agent can transform a raw alert into a structured UAV mission while keeping the process explainable and controlled.
What we learned
The biggest lesson is that the hardest part is not only object detection or drone control. The real challenge is connecting perception, reasoning, safety, and action into one reliable loop.
We learned that an agentic system for physical-world tasks must be:
- Fast, because response time matters.
- Grounded, because every decision must come from real sensor evidence.
- Safe, because UAV actions can affect people and infrastructure.
- Auditable, because every mission needs logs, evidence, and a clear reason.
- Human-supervised, because high-impact actions should not be fully automated.
This project helped us understand how agentic AI can move beyond chat interfaces and become a coordination layer for real-world systems.
Challenges we faced
The main challenge was designing a system that is powerful but still safe.
A UAV response agent cannot behave like a normal chatbot. It must handle uncertainty, missing information, wrong detections, network delay, legal constraints, and physical safety risks. Because of that, we designed the system with verification steps, rule-based guardrails, and human approval for critical actions.
Another challenge was balancing ambition and feasibility. A full real-world deployment would require hardware integration, drone testing, site permission, and legal compliance. For this build, we focused on a clear prototype architecture and mission workflow that can later be connected to real UAV hardware, camera APIs, and edge AI devices.
Why it matters
AeroGuard is designed for environments where every second matters: water safety, renewable energy sites, industrial zones, logistics yards, warehouses, ports, and public safety monitoring.
Many organizations already have cameras. What they lack is an intelligent layer that can decide what happened, what should be checked, which drone should move, and what action should be taken next.
Our vision is to turn passive monitoring systems into active response systems — not by replacing humans, but by helping them verify incidents faster, respond safer, and make better decisions under pressure.
Log in or sign up for Devpost to join the conversation.