sohail-kustagi/MAAS-Disaster-Response

MAAS Disaster Response Models

This repository contains the heavily fine-tuned computer vision and LLM reasoning models used in the Multi-Disaster Autonomous Aerial Swarm (MAAS) project.

Models Included

1. YOLOv10 (Disaster Vision Analyst)

• Architecture: YOLOv10 Nano (best.pt)
• Format: PyTorch and Intel OpenVINO (best_openvino_model/)
• Use Case: Real-time aerial detection of fires, floods, and stranded persons via WebRTC drone streams.
• Training: Fine-tuned over several weeks on a strictly curated disaster dataset to ensure high confidence and minimal false positives in extreme environments.

2. Phi-3-Mini (Disaster Reasoning Commander)

• Architecture: Microsoft Phi-3-Mini 4k Instruct
• Format: GGUF (phi3-lora.gguf) & Safetensors LoRA adapters
• Use Case: Operates as the autonomous reasoning agent. It intakes telemetry data and the YOLO vision detections to autonomously generate MAVLink navigation coordinates without human intervention.

Benchmark Results

We evaluated the models on real-world disaster footage (e.g., Pella Home Fire drone footage).

• Vision Speed (YOLO OpenVINO): 10.33 FPS on edge hardware
• LLM Reasoning Latency: 29.86s per MAVLink command generation

Performance Benchmark Charts

(For the live interactive web demo, please see the linked Hugging Face Space on my profile.)