Drone Swarm Digital Twin
This page must contain only text explicitly present in the provided PDFs. Any missing detail is marked.
Overview
I developed a comprehensive computational framework for autonomous multi-agent systems, combining a modular optimization toolbox with a biologically-inspired drone swarm simulation. The system uses genetic algorithms to evolve decentralized control strategies that produce emergent collective behavior, demonstrating how simple local interaction rules can generate complex mission-level performance without centralized coordination.
The toolbox supports constraint handling, elitism preservation, and parallel fitness evaluation across multiple CPU cores. It includes visualization utilities for convergence analysis, population diversity tracking, and real-time optimization metrics. The architecture is designed for extensibility, new physics engines, cost functions, and constraint handlers integrate through well-defined abstract base classes, making it suitable for applications ranging from trajectory optimization to structural design.
The mission ends only when all targets are captured or all drones are destroyed; surviving agents then autonomously return to base using the same control laws, demonstrating genuine emergent homing behavior without hard-coded return paths.
What I Built / Did
- I created an optimization toolbox that provides a reusable infrastructure for derivative-free optimization problems. At its core is a flexible genetic algorithm implementation with customizable crossover operators (including Zohdi's phi-psi blending for continuous design spaces), tournament and roulette selection mechanisms, and adaptive mutation strategies.
- My emergent swarm control project applies this toolbox to the cooperative target acquisition problem. Instead of optimizing fixed waypoint trajectories, the system evolves control law weights that govern how each drone responds to its local environment.
- Each agent computes its movement force as a weighted sum of four behavioral components: attraction to the nearest unvisited target, repulsion from obstacles, separation from neighboring drones, and cohesion with the swarm centroid.
- The simulation implements realistic physics with configurable timesteps, 3D obstacle fields, and multi-target environments.
- Drones sense their surroundings in real time, identifying closest targets via vectorized distance calculations, detecting obstacles within collision radii using spatial indexing, and maintaining awareness of nearby swarm members for separation and cohesion forces.
- Collision detection runs at every timestep after an initial safe-zone period near the spawn base, with exact intersection criteria for obstacle and drone-drone impacts.
- To achieve computational efficiency at scale, the fitness evaluation is parallelized across all available CPU cores.
- Early generations rely on a fast heuristic cost function (rewarding obstacle avoidance and separation while penalizing cohesion and clustering) to prune the design space; later generations validate top candidates through full physics-based simulation.
- Visualization is handled through a custom 3D animation pipeline built on Matplotlib's `FuncAnimation` framework.