UAV-Adapted Congestion Control

Updated 14 December 2025

UAV-adapted congestion control algorithms optimize data flow by dynamically adjusting transmission thresholds to manage interference and queueing losses.
Cross-layer protocols incorporating altitude-adaptive delay targeting and telemetry headroom reservation ensure low latency and robust command/control transmission.
Reinforcement learning and Markov-based models enhance system resilience to environmental uncertainties and support real-time urban traffic signal management.

A UAV-adapted congestion control algorithm refers to a class of methodologies and protocol refinements specifically tailored for unmanned aerial vehicle (UAV) systems, where the unique constraints of aerial deployment—including wireless link intermittency, mobility-induced loss/latency, interference, onboard resource limitations, and mission-critical packet priorities—require adaptations beyond classical congestion control schemes. This article synthesizes recent advances, including queuing theory for distributed data-plane regulation in UAV mesh networks, cross-layer delay control for video/telemetry transmission, traffic flow stabilization under environment uncertainty, and reinforcement learning for urban congestion mitigation using UAVs as instrumental sensors and actuators.

1. Interference- and Queue-Aware Transmission Control in UAV Networks

Interference-aware congestion control in UAV wireless networks aims to optimize expected throughput under both queueing constraints and spectral interference. The primary technical framework, introduced in "Interference-Aware Queuing Analysis for Distributed Transmission Control in UAV Networks" (Ghazikor et al., 20 Jan 2024), models each node as having an M/M/1 queue with finite buffer capacity $B_n$ and a packet arrival process of intensity $\lambda_n$ . Transmission is regulated by a threshold policy, controlling the probability of sending under current channel conditions, with each packet subject either to buffer overflow or deadline expiration losses.

Analytical Model and Loss Probabilities

Delay Loss: For each node $n$ , the probability a packet exceeds its delay threshold $T_n^{th}$ is

$P_n^{dly}(\beta_n) = \exp\left[ - (\mu_n(\beta_n) - \lambda_n) T_n^{th} \right]$

with $\mu_n(\beta_n)$ the effective slot service probability dependent on threshold $\beta_n$ .

Buffer Overflow: The probability of buffer overflow is approximated via Erlang loss:

$P_n^{ov}(\beta_n) \approx \frac{[1-\rho_n(\beta_n)]\,e^{-B_n\eta_n[1-\rho_n(\beta_n)]}}{1-\rho_n(\beta_n) e^{-B_n\eta_n[1-\rho_n(\beta_n)]}}$

where $\rho_n$ is the offered load.

Interference-induced Outage Probability: Outage due to low SINR is modeled as

$P_n^{out}(\boldsymbol\beta) = \int_{\beta_n}^\infty f_{\tilde h}(x)\, \Pr\left\{ I_n^f > \frac{P_n \hat h_n^2 x^2}{\gamma_{th} - \sigma^2} \right\}\, dx$

with in-band interference $I_n^f$ modeled by a Gamma distribution.

Expected Throughput: Aggregate slot-wise throughput is

$R_n(\boldsymbol\beta) = \lambda_n [1 - P_n^{ov}(\beta_n) - P_n^{dly}(\beta_n) - P_n^{out}(\boldsymbol\beta)]$

Algorithmic Approaches

Interference-Aware Transmission Control (IA-TC): A coordinate-descent algorithm optimizing channel selection thresholds $\boldsymbol\beta$ per link, forcing competing sessions to be conservative under high interference conditions, then tuning the subject link for best trade-off between queuing and channel loss.

Interference-Aware Distributed Transmission Control (IA-DTC): A distributed consensus method, where each node operates as both actor and neighbor. Each runs local search, shares its optimal action, and the procedure iterates toward a Pareto-optimal $\boldsymbol\beta^\star$ vector, maintaining fairness and maximizing network-wide expected throughput.

Tuning Guidelines

Arrival rate $\lambda_n \uparrow$ : Decrease $\beta_n$ , transmit more aggressively to avoid queue drops.
Buffer size $B_n \uparrow$ : Raise $\beta_n$ , allow more waiting for better channel conditions.
Interferer density $N \uparrow$ or SINR threshold $\gamma_{th} \uparrow$ : Increase $\beta$ , transmit more selectively.

2. Cross-Layer Congestion Control for UAV Multimedia and Telemetry

The AQUILA architecture integrates a UAV-adapted congestion control algorithm, SCReAM-FPV, extending the delay-based SCReAM (RFC 8298) paradigm with two critical adaptations: altitude-adaptive delay targeting and telemetry headroom reservation (Huang et al., 7 Dec 2025).

Protocol Enhancements

Altitude-Adaptive Delay Targeting: Rather than fixing a queueing-delay setpoint, the target delay is dynamically scaled to $d_{\rm target}(t) = \max(d_{\rm base},\, \kappa\, \min_{k \in [t-W, t]} RTT(k))$ , utilizing the minimum RTT observed over a window $W$ to absorb handover-induced spikes.
Telemetry Headroom Reservation: Command-and-control (C2) traffic (e.g., MAVLink) is allocated a statically reserved rate $R_{\rm safe}$ , subtracted from the instantaneous transmission rate $R_{tx}$ . The remnant is sent to the video encoder, and any unused "credit" is banked for latency-sensitive packets.

Stepwise Logic

On ACK receipt, compute queueing delay estimate $\hat d_q(t) = RTT(t) - RTT_{\min}(t)$ .
If $\hat d_q(t) \leq d_{\rm target}(t)$ : increase congestion window (additive).
Otherwise: decrease window multiplicatively based on the delay overshot fraction.
Set encoder rate $R_{enc}$ subject to bounds, damping, and reserved headroom.
Credit is accumulated to let C2 packets bypass pacing and maintain bounded latency.

Performance Summary

SCReAM-FPV yields:
- Video one-way latency as low as 213 ms and VMAF 92.20 on LTE traces (vanilla SCReAM: 484 ms, VMAF 56.68).
- PLR for C2 traffic $\approx 0$ throughout; C2 latency ≤ 150 ms, even during link saturation.
- Reconnection time after handover (QUIC 0-RTT) is halved versus TCP/TLS.
- Survives severe bottlenecks (1 Mbps/20 ms): VMAF 39.85, while standard protocols collapse.

3. Resilient UAV Traffic Congestion Control under Environmental Uncertainty

Traffic congestion control for UAV fleets in the airspace leverages continuous-time fluid queuing models, where link capacities fluctuate as a Markov chain driven by weather uncertainty (Zhou et al., 2019). This aligns capacity allocation and inflow regulation for single-point, tandem, and merge linkage topologies.

Mathematical Formulation

Queue Dynamics: For a queue with inflow $a$ and mode-dependent capacity $c_i$ :

$\dot Q(t) = a - f(I(t), Q(t)),\quad f(i, q) = \begin{cases} c_i, & q > 0 \ \min(c_i, a), & q=0 \end{cases}$

Markov Chain: Weather state $I(t)$ follows a CTMC with generator $Q = (\lambda_{ij})$ .
Stability (Resilience) Condition: Long-term stochastic stability if

$\sum_{i=1}^m p_i c_i \ge a$

for a single queue, or for tandem/merge links, analogous aggregate inequalities.

Throughput Optimization: The largest inflow $a_s$ guaranteeing stability can be found via bisection on feasible bilinear Lyapunov inequalities.

Real-Time Congestion Control

Discrete-time event-driven pseudocode at each $\Delta t$ :

Measure queue lengths and current mode.
Compute provisional capacities, allocate flows proportionally.
Update queue states and advance CTMC.

Practical Insights

Sufficient stability bound $a_s$ is typically below the necessary average $\bar a$ during sustained adverse weather modes.
Mode-switching intensity $\lambda_{ij}$ "smooths" short-term variance.
Offline cataloguing of safe inflow regimes enables operational enforcement of $a \le a_s$ .

4. UAV-Actuated Urban Traffic Signal Congestion Control

The AVARS system explores UAVs as rapid-response actuators for urban traffic signals, employing deep reinforcement learning (DRL) for congestion alleviation (Guo et al., 2023). UAV-mounted cameras enable high-frequency, high-resolution traffic monitoring at targeted intersections.

MDP and Algorithmic Framework

MDP Structure: Each UAV–traffic intersection pair is modeled as $(S, A, P, R, \gamma)$ $(S, A, P, R, γ)$ .
- State space $S$ : Concatenated current phase, per-lane occupancy $O_i(t)$ , and speed $v_i(t)$ .
- Action space $A$ : Binary, $\{0,1\}$ (hold/switch phase).
- Reward: $r_t = - \max_{i \in I} O_i(t)$ , directly penalizing peak congestion.
UAV Constraints: Battery lifetime as episode horizon ( $B_0$ ); coverage limited to $R_{max}$ .
DRL Methods: DQN (value-based) and PPO (policy-gradient), both with lightweight MLP architectures.

Control Procedure Outline

Pre-train policy networks in simulation.
On event detection, deploy UAVs to highest-impact intersections.
At each time step:
- Observe local state.
- Select/execute action (signal phase switch or hold).
- Update local experience buffer.
- Decrement battery.
Post-episode, aggregate trajectories and update central DRL model.
Repeat as untoward events arise and for continual fine-tuning.

Empirical Results

Metric	Original	Congestion	SCATS	IntelliLight	AVARS
Travel time (s)	275.8	597.3	429.6	232.7	195.4
Fuel cons. (l/100 km)	20.14	45.45	31.69	14.60	12.31
CO $_2$ (g/km)	468.6	1057.2	737.1	339.6	286.3

AVARS achieved up to 73% reductions in delay and emissions, outperforming fixed-time, loop-adaptive, and conventional DRL baselines within regular UAV battery duration.

5. Design Trade-Offs and Tuning Guidelines

UAV-specific congestion control algorithms are characterized by critical trade-offs:

Lowering transmission or control activation thresholds reduces in-band interference or control-induced queueing loss, but risks increased packet drop or reduced control responsiveness.
Buffer sizes permit more tolerance for channel variability, allowing more aggressive congestion control strategies.
Environmental variables (weather, interferer density, SINR thresholds) must be actively measured and incorporated into control adaptation logic.
Offline parameter space exploration (Lyapunov-based feasibility checks) should precede deployment; online algorithms require slot-wise adaptation based on measured arrival rates, delay constraints, and network interference profile (Ghazikor et al., 20 Jan 2024, Zhou et al., 2019).

6. Impact and Future Directions

UAV-adapted congestion control is central to the practical deployment of multi-UAV systems for communication, sensing, cloud offloading, and autonomous mobility. Methodological advances—such as interference-aware queuing, cross-layer delay-based protocols, resilience modeling via Markov fluid queues, and real-time DRL for urban applications—significantly improve both safety-critical link integrity and end-user quality of service. Future directions include:

Integration of UAV congestion control with next-generation cellular and satellite networks.
Joint optimization of energy, spectrum, and data plane for SWaP (size, weight, and power)-limited airborne platforms.
Robust control under adversarial or non-stationary environments.
Extension to coordinated multi-agent UAV traffic management in dense urban or disaster-response scenarios.

These frameworks lay the technical foundation for reliable, scalable UAV operations under realistic wireless and traffic conditions.