Semantic-Aware UAV Command and Control for Efficient IoT Data Collection

1. Problem Statement

The paper addresses the challenge of efficiently collecting image data from Internet of Things (IoT) devices using Unmanned Aerial Vehicles (UAVs) in scenarios with limited communication, energy, and bandwidth resources.

Key Challenges:

• Resource Constraints: UAVs have limited battery and bandwidth, making the transmission of massive raw image data inefficient.
• Trajectory vs. Data Quality Trade-off: The UAV must balance visiting as many devices as possible (coverage) with staying within communication range long enough to transmit sufficient data for high-quality image reconstruction.
• Command and Control (C&C) Latency: Existing approaches often assume ideal, delay-free C&C signals. In reality, there is a delay in transmitting acceleration commands from the Base Station (BS) to the UAV, which complicates real-time trajectory optimization.
• Partial Transmission: Due to mobility, devices may not transmit their full data payload before the UAV leaves the communication range. Traditional coding schemes fail under these conditions, whereas semantic communication offers a solution.

2. Methodology

The authors propose a novel framework integrating Semantic Communication with UAV Command and Control, optimized via Reinforcement Learning (RL).

A. System Model

• Semantic Encoding (DeepJSCC): IoT devices use Deep Joint Source-Channel Coding (DeepJSCC) to encode images into compact, complex-valued latent vectors. Crucially, these vectors are ordered by importance (magnitude). This allows the receiver to reconstruct an image even if only a partial set of symbols is received (by padding missing symbols with zeros). The more symbols received, the higher the Peak Signal-to-Noise Ratio (PSNR).
• Communication Architecture:
  • Uplink: IoT devices transmit symbols to the UAV using OFDMA when within range ($D_{com}$).
  • Downlink (C&C): A Base Station (BS) controls the UAV's trajectory by sending 2D acceleration commands ($a[t]$).
  • Delay Modeling: The system explicitly models a downlink delay ($\Delta[t]$) for the C&C signals from the BS to the UAV, reflecting realistic computational and transmission lags.
• Dynamics: The UAV moves in discrete time slots ($K$) with fixed altitude. Its motion is governed by kinematic equations where velocity and position are updated based on acceleration commands.

B. Problem Formulation

The objective is to maximize the average PSNR of reconstructed images across all visited devices within a fixed time horizon $T$.

• Decision Variable: The sequence of acceleration commands ($A$) generated by the BS.
• Constraints:
  • Each device is visited at most once.
  • UAV dynamics (velocity and position updates).
  • Final destination requirement.
  • Acceleration bounds.
• Optimization Goal: Maximize the ratio of collected symbols to total possible symbols, weighted by the semantic importance, effectively maximizing image quality.

C. Proposed Solution: Double Deep Q-Learning (DDQN)

Since the problem involves sequential decision-making under uncertainty (channel conditions, device connectivity, and C&C delays), the authors model it as a Markov Decision Process (MDP) and solve it using Double Deep Q-Learning (DDQN).

• State Space ($S$): Includes UAV position/velocity, visited devices, current image quality metrics, and remaining mission time.
• Action Space ($A$): Discrete 2D acceleration vectors for the next time slot.
• Reward Function ($R$): A composite reward designed to:
  1. Encourage data collection and visiting multiple devices ($r_c$).
  2. Ensure the UAV reaches the final destination within the time limit ($r_d$).
  3. Provide a bonus for successful mission completion ($r_g$).
• Algorithm Features:
  • Uses an Online Network ($\theta$) for action selection and a Target Network ($\theta^-$) for value estimation to reduce overestimation bias.
  • Employs a Replay Buffer to store transitions $(S, a, r, S')$ for stable training.
  • The agent learns an adaptive flight policy that dynamically adjusts speed and hovering time based on the semantic value of the data and the delay in command execution.

3. Key Contributions

1. Semantic-Aware Framework: Integration of DeepJSCC with UAV trajectory optimization, allowing for robust image reconstruction even with partial data transmission.
2. MDP Formulation with Realistic Constraints: The first formulation to explicitly account for downlink C&C delays and the trade-off between trajectory optimization and semantic data recovery quality.
3. Adaptive DDQN Control: Development of a DDQN-based algorithm that learns to balance speed, coverage, and hovering time, outperforming static or greedy strategies.

4. Simulation Results

The framework was tested in a $1000 \times 600$ m area with 10 IoT devices, comparing the DDQN approach against Greedy (nearest neighbor) and Traveling Salesman Problem (TSP) baselines.

• Performance vs. Bandwidth: The DDQN approach achieved higher average PSNR across all bandwidths. At low bandwidths, baselines suffered due to limited symbol transmission, while DDQN adapted by optimizing hovering time to maximize symbol collection.
• Performance vs. Velocity:
  • Low Speed: Baselines failed to visit enough devices.
  • High Speed: Baselines moved too fast, causing incomplete transmissions and poor reconstruction.
  • DDQN: Dynamically adjusted speed, slowing down near devices to ensure high-quality data capture while maintaining mission progress.
• Trajectory Analysis: Unlike the rigid paths of TSP or the short-sightedness of Greedy, the DDQN trajectory was dynamic, allowing the UAV to linger strategically around devices to maximize semantic data recovery.

5. Significance

This work represents a significant step forward in 6G and IoT-enabled UAV networks by shifting the focus from "bit-perfect" transmission to "task-relevant" information recovery.

• Efficiency: It demonstrates that semantic communication can drastically reduce bandwidth requirements while maintaining high utility (image quality).
• Robustness: By accounting for C&C delays, the proposed control policy is more viable for real-world deployment where latency is inevitable.
• Scalability: The use of Deep Reinforcement Learning provides a scalable solution for complex, dynamic environments where traditional optimization methods (like convex optimization) struggle to find global optima in real-time.

In conclusion, the paper proves that integrating semantic coding with intelligent, delay-aware trajectory control significantly enhances the efficiency and quality of IoT data collection in UAV networks.