UAV Path Planning with Deep Q-Networks

Final year project, BEng Electrical Engineering, University of Manchester.

A DQN agent flies a UAV over a 2D grid to collect data from LoRa IoT sensors, balancing throughput, energy, and coverage fairness.

What I built

• Custom Gymnasium environment simulating LoRaWAN physics (Two-Ray path loss, EMA-ADR, AoI)
• DQN with domain randomisation across 16 environment conditions
• Competence-based curriculum learning (NDR >= 95%, Jain fairness >= 0.85)
• 4-frame state stacking with cross-attention sensor encoding
• Full ablation study (A1-A4) with Welch t-tests vs baselines

Results

• 100% sensor coverage, matching the greedy oracle baseline
• 72,590 bytes collected per episode
• Cohen d = 1.43 vs combined ablation
• High first-class grade

Stack

Python 3.11, Stable-Baselines3, PyTorch, Gymnasium, Matplotlib, Seaborn.