The Thesis.
Single drones are limited by endurance and scope. My research focuses on swarm intelligence to offer the resilience and scale necessary for complex real-world operations.
Core OS
Ubuntu 24.04
Framework
ROS2 Jazzy
Simulation
Gazebo Harmonic
Control
MRS UAV System
Learning
PyTorch DQN
Process & Validation.
Utilizing the MRS UAV ROS2 Control Framework, I validated swarm coordination and data integrity within the Gazebo Harmonic environment.
High-fidelity swarm coordination on Ubuntu 24.04
Verified 100Hz odometry updates for decentralized logic
Research Poster.
Full documentation of the autonomous drone swarm project presented at UNC Charlotte Honors College.
RL Coverage Optimization.
Training reinforcement learning agents to solve multi-drone coverage — positioning 3 drones to observe all 25 target points across a 100x100m area. The problem is first abstracted in 2D, then transferred to full 3D Gazebo simulation.
Algorithm
Deep Q-Network
Framework
PyTorch
Baseline
Tabular Q-Learning
Sim Bridge
Gazebo SDF World
25 dots on a 5x5 grid — disks represent drone coverage footprints (r=1.2)
Experience replay, target network updates, epsilon-greedy exploration with decay
The DQN agent uses a 4-layer fully connected network (4 → 64 → 64 → 64 → 25) with experience replay and a target network updated every 1,000 steps. When a convergence plateau appeared, the task was reframed as a static placement problem rather than a sequential decision process — a grid-search optimizer then established a 76% coverage benchmark the RL agent trains against. The companion DiscWorld repo generates Gazebo SDF worlds with 25 cylindrical targets for sim-to-real transfer.
Research Output.
This work is being written up for the AIAA SciTech 2027 Student Paper Competition (Uncrewed and Autonomous Systems track), co-authored at SmartNet Lab under Adv. Dr. Ran Zhang. The pipeline — 2D Q-learning prototype → DQN with benchmark → full ROS2/Gazebo simulation — documents the complete iteration arc from abstraction to deployment-ready system.
Repositories.
The full research arc lives across four repos — from a 50-line NumPy prototype to a production ROS2 simulation stack.
The disk-position prototype has its own page documenting the RL formulation and the lineage from 2D abstraction to 3D Gazebo deployment.