Abstract
Operating Unmanned Aerial Vehicle (UAV) swarms in dynamic environments necessitates high-precision trajectory prediction and efficient collision avoidance. This paper presents a novel high-precision AI framework featuring the AdaptoSwelliGauss activation function and the Integrated Collision Detection, Avoidance, and Batching (ICDAB) strategy. The proposed Feedforward Neural Network (FFNN) achieves trajectory prediction errors significantly lower than existing benchmarks, while the ICDAB method optimizes swarm takeoff efficiency by balancing trajectory adjustments with batching.
Introduction
The deployment of UAV swarms in complex scenarios requires real-time trajectory adjustments to maintain safety and mission objectives. Traditional non-linear optimization techniques often lack the speed required for real-time applications. While Feedforward Neural Networks (FFNNs) offer a faster alternative, their accuracy has historically been limited. Previous research typically reported Mean Squared Errors (MSE) in the $10^{-3}$ range, which may be insufficient for high-density swarm operations.
Furthermore, collision avoidance remains a critical bottleneck. Existing methods often rely on either complex path re-planning, which can lead to inefficient flight paths, or simple batching (delayed takeoff), which increases total mission time. This research seeks to bridge the accuracy gap in trajectory prediction and provide a more efficient middle ground for collision management.
Methodology
Neural Network Architecture
The researchers implemented a Feedforward Neural Network (FFNN) designed to predict future waypoints based on initial, middle, and final states. The architecture consists of an input layer representing these states, a single hidden layer with 15 neurons, and an output layer with 201 neurons corresponding to the trajectory waypoints.
AdaptoSwelliGauss Activation Function
A primary innovation of this work is the AdaptoSwelliGauss activation function. Recognizing that UAV flight involves both smooth gliding and abrupt maneuvers, the authors developed a piecewise hybrid function that switches components based on a hyperparameter $\alpha$:
- Swish: Ensures smooth transitions for trajectory continuity.
- Elliott: Captures rapid changes in velocity or direction.
- Scaled/Shifted Gaussian: Enhances robustness against data noise.
ICDAB Strategy
To address collision avoidance, the Integrated Collision Detection, Avoidance, and Batching (ICDAB) strategy was introduced. It builds upon the Circular Arc Trajectory Geometric Method (CTGA) but introduces a constraint to prevent infinite optimization loops. The algorithm limits trajectory manipulations to 10 attempts per UAV. If a collision remains unresolved after these attempts, the UAV is assigned to a batching list for a delayed takeoff, ensuring both safety and computational convergence.
Results and Discussion
Prediction Accuracy
The AdaptoSwelliGauss function demonstrated superior performance compared to standard activation functions like ReLU and Tanh. In predicting the X-coordinate of trajectories, the proposed model achieved an unprecedented MSE of $3.059 \times 10^{-14}$. This represents a multi-order of magnitude improvement over ReLU ($1.733 \times 10^{-10}$) and Tanh ($1.406 \times 10^{-7}$).
Swarm Efficiency
The ICDAB algorithm significantly improved swarm management. In a simulation involving 262 potential initial collisions, the strategy successfully resolved the majority, leaving only 33 UAVs to be handled via batching. This integrated approach reduced the total number of takeoff batches from 7 to 4. Additionally, the system increased the maximum number of UAVs capable of simultaneous takeoff in a single batch from 38 to 396, drastically enhancing mission throughput.
Conclusion
This study introduces a robust framework for UAV swarm trajectory management. By combining a specialized hybrid activation function with a constrained collision avoidance strategy, the researchers achieved high-precision predictions and improved operational efficiency. Future work will focus on validating these results with real-world sensor data and exploring more complex architectures like Multi-Agent Reinforcement Learning (MARL) for collaborative swarm behavior.
Summary updated