Overview
Drone AI systems represent one of the most demanding applications of embodied artificial intelligence: decisions must be made in milliseconds, compute budgets are measured in watts, and failures carry physical consequences. At AppSofa Lab, we research the full stack of AI capabilities required for autonomous drone operation — from low-level flight control to high-level mission planning.
Our work spans both the software and hardware integration challenges of putting modern AI models on aerial platforms, with applications in federal reconnaissance, infrastructure inspection, and autonomous delivery.
Autonomous Navigation and SLAM
Simultaneous Localization and Mapping (SLAM) enables a drone to build a map of an unknown environment while simultaneously tracking its own position within it — all in real-time, without GPS.
GPS-denied environments — indoor facilities, urban canyons, underground structures — are critical for both federal and commercial drone applications. Our navigation research focuses on robust SLAM pipelines that fuse multiple sensor modalities:
Visual-Inertial Odometry (VIO)
Combines camera frames with IMU data for accurate 6-DOF pose estimation at low computational cost.
LiDAR-based SLAM
Point cloud registration for centimeter-level accuracy in structured environments.
Neural SLAM
Learning-based scene representations (NeRF, 3D Gaussian Splatting) for real-time dense mapping with semantic understanding.
Path Planning
Online replanning algorithms — RRT*, trajectory optimization — for dynamic obstacle avoidance in unstructured environments.
Real-Time Computer Vision
Aerial platforms require perception systems that are simultaneously fast, accurate, and power-efficient. Our computer vision research targets drone-specific challenges — small object detection at high altitude, motion blur, changing illumination, and cluttered backgrounds.
Object Detection
Real-time detection of vehicles, personnel, and infrastructure at varying altitudes using lightweight YOLO-family and transformer-based detectors.
Semantic Segmentation
Pixel-level scene understanding for terrain classification, landing zone identification, and obstacle mapping.
Target Tracking
Multi-object tracking (MOT) with re-identification across frames and viewpoints, robust to occlusion and fast motion.
Change Detection
Comparing aerial captures across time to identify structural changes, anomalies, or activity patterns.
Swarm Intelligence
Single drones are limited by sensor range, battery life, and fault tolerance. Swarms distribute these constraints across many cooperating agents. Our swarm research addresses the coordination algorithms that enable coherent collective behavior to emerge from decentralized local decisions:
- Decentralized task allocation — Distributed assignment of missions across swarm members without a central coordinator — robust to communication dropout and individual failures.
- Formation control — Maintaining geometric configurations during transit while adapting dynamically to obstacles and terrain.
- Collaborative perception — Fusing observations from multiple UAVs to build a shared situational awareness map with higher coverage and lower latency than any single platform.
- Adversarial robustness — Designing swarm behaviors resilient to jamming, spoofing, and adversarial interference — critical for federal deployment.
Edge AI Deployment
Models trained on our NVIDIA GPU clusters must be compressed and optimized for deployment on embedded platforms with strict SWaP (Size, Weight, and Power) constraints. Our edge AI work bridges the gap from research to field hardware:
Collaborate
Working on autonomous systems or UAV applications?
We partner with federal agencies and commercial operators on autonomous drone AI — from research prototypes to field-deployable systems.
Get in Touch