Wildfire
Surveillance
Autonomy
From Fast-Mixing Coverage to Smooth Risk-Aware Aerial Search
Wildfire surveillance is not one control problem but a stack of them. This project connects stochastic coverage, scalable Markov-policy design, continuous-time motion design, and distributed quadrotor surveillance into a layered autonomy framework for large uncertain environments.
A Connected Multi-Layer Research Program
This page synthesizes a PhD research program spanning stochastic coverage, scalable optimization, continuous-time control, and distributed coordination into one coherent autonomy stack. The layout follows the dedicated wildfire baseline from v4, while the data and related research links come from the current site as the source of truth.
Regional surveillance requires allocating attention over a changing risk field. That allocation must remain computable as scale grows. Local search must be smooth and dynamically feasible. Distributed agents must operate under local sensing and communication limits.
The project treats these as connected layers of the same challenge: how do you make a drone swarm survey a large unknown environment efficiently, smoothly, and with provable behavior?
Why This Problem Is Hard
Six distinct technical challenges motivate the layered approach. Solving any one in isolation is insufficient.
Wildfire risk covers tens of thousands of km2 with a spatially complex, nonstationary utility distribution. No single agent can observe or plan globally.
Risk fields modeled as Gaussian Mixture Models have multiple local maxima and plateau regions where naive gradient ascent stalls completely.
Global Markov-chain design over a 100x100 region map produces 10^4 variables. Without decomposition, optimization time scales poorly with map size.
Each drone can observe only local utility and neighbors within a limited communication radius. There is no access to a global gradient or full state.
Quadrotor dynamics require smooth, trackable references. Discrete optimization outputs or discontinuous switching produce physically infeasible trajectories.
Agents must explore cooperatively without redundant re-entry into already-covered regions while relying only on intermittent local communication.
System Architecture
Four connected layers together constitute a full autonomy stack, from high-level allocation to executable flight behavior.
Risk-Aware Allocation
How surveillance effort is distributed over a partitioned region graph using Markov-based guidance.
The terrain is divided into a region graph. A Markov transition matrix is designed to produce the fastest-mixing chain to a desired stationary distribution proportional to wildfire risk. Minimizing the second-largest eigenvalue modulus guides the swarm stochastically toward high-risk regions while preserving graph constraints.
Scalable Policy Computation
How the Markov design problem is decomposed into bounded local SDPs that scale to larger maps.
Naive global Markov design becomes computationally prohibitive as the map grows. The Smallest-SDP decomposition exploits sparse local motion, reducing the problem into bounded local blocks with overlap coordination and ADMM-style stitching.
Each local SDP block stays small enough to keep the global design tractable as the map grows.
Continuous-Time Motion Design
How accelerated optimization ideas become smooth motion logic for drone path planning.
Discrete optimization updates are too jagged for real vehicle dynamics. Continuous-time analogues of Heavy-Ball, Nesterov, and Triple-Momentum methods yield smoother reference trajectories. Hybrid switching keeps progress through nonconvex fields without sacrificing trackability.
Hybrid switching escapes plateau regions while preserving smooth, dynamically feasible motion.
Distributed Executable Surveillance
How local information, flocking, spiral coverage, and stabilization become real swarm behavior.
Each drone acts on local utility and neighbor communication only. A gradient-free surrogate, flocking forces, and hybrid mission logic drive search behavior, while ellipse fitting and spiral coverage convert detection into contiguous basin inspection.
Local decisions compose into global coverage without assuming a centralized global gradient.
Evidence
These result blocks summarize the strongest evidence across the main layers of the wildfire stack.
Bounded Local Blocks
Scalability came from the decomposition structure, not just from choosing a different solver.
The Smallest-SDP method exploits sparse local movement structure and decomposes the global optimization into bounded local blocks, which shifts the growth of computation toward a much more practical regime as maps scale.
Local structure keeps the policy-design subproblems small as the mission map grows.
Hybrid Controller for Full Coverage
The contribution is not only smooth trajectories, but executable surveillance behavior under realistic motion constraints.
Hybrid switching combines global ascent and local refinement so the quadrotor can keep making progress through plateau regions while remaining smooth enough to track with nonlinear control.
Smoothness is treated as a control requirement, not a visual nicety.
Dual-Flock Coordination
Shared discovery memory materially improves distributed search efficiency.
Coordinated flocks that exchange discovered basin information avoid redundant re-entry and shorten the overall mission compared with a single-flock configuration.
Distributed execution improves when local discoveries become shared mission memory.
10,000 km2 Portugal Map
The large-scale validation is useful because it exposes the framework's boundary conditions instead of hiding them.
The Portugal wildfire-risk scenario makes charging, endurance, maneuverability, and local-gradient limitations concrete while still showing meaningful coverage of high-risk regions.
Large-scale validation makes the endurance and maneuverability tradeoffs explicit.
My Role
This was a PhD research project. The portfolio framing here emphasizes system design, implementation, and how the research layers connect into an executable autonomy stack.
Framed wildfire surveillance as a connected multi-layer autonomy problem spanning allocation, policy computation, motion design, and distributed execution.
Designed the Markov-chain optimization formulation and SDP decomposition strategy for scalable policy computation.
Developed continuous-time gradient methods with optimized parameters and derived implementation-oriented analysis.
Formulated the hybrid optimization framework for nonconvex utility surveillance with smooth trackable motion.
Designed the distributed gradient-free flocking and hybrid surveillance controller for multi-agent swarms.
Built the simulation and validation narrative across the full stack, including the Portugal wildfire-risk practical scenario.
Tradeoffs & Technical Decisions
Engineering judgment, not just mathematical results. These are the decisions that shaped what this system actually is.
Decomposition Before Solver Choice
The scalability gain came from problem structure, not from picking a different optimizer on the monolithic formulation.
Smoothness as a Hard Requirement
Drone trajectories had to remain trackable under vehicle dynamics, which made continuous-time motion design a first-class design requirement.
Local Information Instead of Full State
The system is intentionally constrained to local sensing and neighbor communication rather than an idealized centralized view.
Hybrid Switching for Nonconvex Landscapes
Pure gradient ascent cannot reliably escape plateaus, so hybrid switching became central to the surveillance behavior rather than an optimization detail.
Shared Discovery Memory for Coordination
Without memory of discovered optima, multi-flock search wastes mission time by re-entering already-covered basins.
Explicit Endurance Tradeoffs
Charging-stop and fixed-wing comparisons were part of the design story, helping make maneuverability and endurance tradeoffs concrete.
Limitations & Honest Assessment
These boundaries make the results more credible. The system is strong where it has evidence, and explicit about where future extensions are still needed.
The gradient-free surrogate can be dominated by nearby steeper basins, which makes tightly clustered optima harder to separate cleanly.
The current framework assumes fixed-altitude operation and an initially connected flock.
The surrogate incurs approximation error that trades off against inter-agent spacing and exploration breadth.
Mission time still depends non-trivially on actuator limits, local curvature, and formation size even when the asymptotic theory is sound.
Connected Outputs
The dedicated v4-style wildfire narrative is paired here with the current site's richer supporting material: related publications, technical writing, and supporting project infrastructure.
Smooth surveillance using quadrotors for tasks with nonconvex utility functions
Systems & Control Letters
PublicationContinuous trajectory planning for non-convex utility functions using hybrid optimization
European Journal of Control
PublicationA self-organizing distributed algorithm to tackle the stochastic coverage problem
Franklin Open
PublicationAnalysis of gradient descent algorithms: Discrete to continuous domains and circuit equivalents
Systems & Control Letters
PublicationDistributed Surveillance System with Drone Formations
IEEE Transactions on Control of Network Systems
Designing the Markov-Matrix: Why Adapting a Random Walk Fell Short
An early note on Markov chains, swarm density, and why mixing speed became the real problem.
WritingDynamics of Optimization: From Faster Iterations to Faster Settling
An early note on continuous optimization, circuit equivalents, and why settling time became the real design problem.
WritingQuadrotor Surveillance: Local Peak Seeking and Smooth Coverage Control
An early note on unknown utility maps, hybrid control, and why smooth reference generation became the real problem.
WritingMarkov Matrix Design: From Global SDP to Local ADMM
How a decomposition-first view makes large stochastic-coverage optimization practical.
Fast-Mixing Policies for Stochastic Coverage
research thread
LabLocal ADMM for Markov-Matrix Design
research thread
LabSettling-Time Design for Accelerated Optimization Flows
research thread
LabCircuit Equivalents of Optimization Dynamics
research thread
LabSmooth Coverage over Unknown Utility Fields
research thread
LabRestart Logic for Multi-Peak Aerial Search
research thread
LabFormation-Aware Distributed Surveillance
research thread