💻 computer science

Decentralized and Fully Onboard: Range-Aided Cooperative Localization and Navigation on Micro Aerial Vehicles

This paper presents a fully decentralized framework for micro aerial vehicles that combines onboard odometry and inter-robot range measurements with a block coordinate descent localization algorithm and a factor graph-based formation control approach to achieve accurate, coordinated navigation without relying on centralized computation or external global positioning systems.

Original authors: Abhishek Goudar, Angela P. Schoellig

Published 2026-02-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abhishek Goudar, Angela P. Schoellig

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are leading a group of friends through a dense, foggy forest. You don't have a map, and you can't see the stars (no GPS). In the past, you might have needed a giant, expensive tower outside the forest to shout directions to everyone. But what if that tower doesn't exist, or is too expensive to build?

This paper presents a clever solution for Micro Aerial Vehicles (MAVs)—basically, tiny, smart drones—to fly together in a perfect formation without needing any outside help.

Here is the breakdown of how they do it, using some everyday analogies:

1. The Problem: The "Blind" Team

Usually, robots rely on two things to know where they are:

GPS: Like a lighthouse guiding a ship. (But it doesn't work indoors or under thick trees).
Central Brain: One big computer telling everyone where to go. (But if that computer crashes, the whole team crashes).

The authors wanted a team of drones that could fly together, stay in a specific shape (like a triangle), and know exactly where they are, even if they are in a metal dome or a forest with no signal. They needed to be decentralized, meaning every drone is its own boss, but they all cooperate.

2. The Solution: The "Blind Man's Bluff" with a Twist

Instead of looking at the world, the drones look at each other.

The Sensors: Each drone has a camera, an accelerometer (to feel movement), and a "ruler" (a radio that measures distance to neighbors).
The Magic Trick (Localization): Imagine you are blindfolded in a room with two friends. You can't see, but you can feel how much you walked (odometry) and you can ask your friends, "How far away are you?"
- If you walk forward and your friend says, "You are now 2 meters closer," you can figure out your position relative to them.
- The paper uses a mathematical trick called Factor Graphs. Think of this as a giant, invisible web of rubber bands connecting the drones. Every time a drone moves or measures a distance, it pulls on the rubber band. The system calculates the tension to figure out exactly where everyone is, even if the measurements are a little "noisy" or fuzzy.

3. The "Smart" Formation Control

Usually, keeping a formation is like a dance where everyone follows a strict script. If the music skips (sensor noise), the dancers trip.

The authors made the dance adaptive:

The Metaphor: Imagine the drones are holding hands with elastic bands. If one drone gets a little shaky or unsure of its position (high uncertainty), the "elastic band" to its neighbor gets looser. The system says, "Okay, we aren't 100% sure where you are, so let's not panic if we drift a tiny bit."
The Result: Instead of a rigid, brittle formation that breaks easily, they have a flexible, resilient formation. It accounts for the fact that their sensors aren't perfect, so they don't overreact to small errors. This keeps the formation smooth and stable.

4. How They Talk (The "Asynchronous" Chat)

In many robot teams, everyone has to talk at the exact same millisecond. If one drone is slow, the whole team waits. That's like a group chat where everyone has to reply instantly or the conversation stops.

This system is asynchronous.

The Metaphor: It's like a group text message. You send a message, and your friends reply whenever they can. You don't wait for everyone to reply before you send your next thought.
Why it helps: If a drone loses a signal for a second or is busy calculating, the others keep moving. They just use the "last known good info" until the laggy drone catches up. This makes the system robust against bad Wi-Fi or slow computers.

5. The Real-World Test

The researchers didn't just simulate this on a computer; they flew real drones.

The Test: They flew three drones in a triangle inside a giant metal dome (where GPS is impossible) and under a tree canopy (where wind and leaves mess up cameras).
The Outcome: The drones stayed in formation with decimeter-level accuracy (within about 4 inches of the perfect spot). They did this without a central computer, without GPS, and without needing to talk in perfect sync.

Summary

Think of this paper as teaching a flock of birds how to fly in a perfect V-shape during a storm, without a leader bird calling out orders. Instead, every bird just watches its neighbors, feels the wind, and adjusts its wings based on how confident it is about its own position. If the wind gets crazy, the flock loosens up slightly but never breaks apart.

The takeaway: You don't need a super-computer or a GPS tower to coordinate a team of robots. You just need them to trust each other, communicate flexibly, and be smart about their own uncertainty.

1. Problem Statement

The paper addresses the challenge of controlling a team of Micro Aerial Vehicles (MAVs) to fly in a coordinated formation without relying on centralized computing or external infrastructure (like GPS or pre-installed UWB anchor networks).

Key Challenges:

Scalability & Robustness: Centralized approaches suffer from single points of failure and poor scalability.
Infrastructure Limitations: GPS is unavailable indoors or in dense urban canopies; visual landmarks require persistent features and are computationally expensive.
Sensor Noise & Uncertainty: Off-the-shelf range sensors (e.g., UWB) are noisy and prone to non-line-of-sight (NLOS) errors. Traditional formation control often ignores the uncertainty in state estimation, leading to poor tracking performance.
Communication Constraints: Real-world networks suffer from latency, packet loss, and asynchronous data arrival, making strict synchronization difficult.

Objective:
To develop a fully decentralized system where each MAV estimates its own pose and calculates control inputs to maintain a desired distance-based formation using only onboard sensors (Visual-Inertial Odometry, IMU, and Range sensors) and local communication with neighbors.

2. Methodology

The authors propose a two-stage framework: Decentralized Cooperative Localization and Decentralized Formation Control, both modeled as inference problems on factor graphs.

A. Decentralized Cooperative Localization

The goal is to estimate the pose ( $x \in SE(d)$ ) of each robot using a combination of Visual-Inertial Odometry (VIO) and inter-agent range measurements.

Factor Graph Formulation: The joint probability distribution of robot states and measurements is modeled.
- Odometry Factors: VIO pose estimates are pre-integrated to reduce graph size, creating relative pose constraints between time steps.
- Range Factors: Distance measurements between agents provide constraints linking the states of different robots.
Optimization Approach:
- The Maximum A Posteriori (MAP) estimation is formulated as a non-linear least squares problem.
- Asynchronous Block Coordinate Descent (BCD): Instead of solving the global optimization simultaneously, each agent iteratively solves a local sub-problem using the latest state estimates from its neighbors.
- Uncertainty Handling: The range factor is modified to account for the uncertainty of the neighbor's state. If a neighbor's state has high variance, the corresponding distance constraint is weighted less (penalized more) in the optimization.
- Asynchrony: The system handles network delays and packet losses by adding delayed measurements to the factor graph at their correct historical timestamps or removing them if data is missing. A fixed-lag smoother is used to keep computational costs low.

B. Decentralized Formation Control

The goal is to generate smooth control inputs to maintain a rigid, distance-based formation while tracking a leader's trajectory.

Inference-Based Control: Formation control is framed as a MAP inference problem on a factor graph, distinct from traditional rigid-body graph theory approaches.
Motion Priors: A continuous-time Gaussian Process (GP) regression is used to generate smooth motion priors. This incorporates the agent's own motion model uncertainty, ensuring that if estimation uncertainty is high, the generated trajectory remains smooth.
Distance Constraints: Desired inter-agent distances are incorporated as soft constraints (factors) in the cost function rather than hard equality constraints. This allows the system to relax formation requirements temporarily if communication is unreliable or obstacles are encountered.
Uncertainty Propagation: The cost function explicitly includes the marginal uncertainty of neighboring agents. If a neighbor's position is uncertain, the controller adjusts its inputs to be more conservative regarding that specific constraint.
Execution: Each agent independently solves the optimization over a finite time horizon using a fixed-lag smoother, outputting position or velocity setpoints for the low-level flight controller.

3. Key Contributions

Novel Inference-Based Formation Control: A new formulation of formation control as a sparse inference problem on factor graphs that explicitly accounts for state estimation uncertainty and sensor noise, leading to smoother and more robust control.
Decentralized Asynchronous Localization: A cooperative localization approach using range, visual, and inertial sensors that utilizes asynchronous Block Coordinate Descent. It does not require synchronized communication or the exchange of raw odometry data, only pose estimates and covariances.
Real-World Validation: Successful demonstration of joint localization and formation control on resource-constrained MAVs in diverse environments (indoor, outdoor, GPS-denied, and under tree canopies) without specialized trajectories.

4. Results

The approach was evaluated through simulations and 22 real-world flight experiments involving up to 3 MAVs.

Localization Accuracy:
- The proposed Decentralized Cooperative Localization (DCL) achieved positioning Root Mean Square Error (RMSE) comparable to a centralized batch trajectory estimation baseline.
- In real experiments with UWB precision of $\pm 0.1$ m, the DCL method achieved sub-decimeter accuracy (e.g., $\approx 0.10$ m RMSE).
Formation Control Performance:
- The proposed method significantly outperformed a baseline gradient control method (GCM). The baseline suffered from instability due to communication delays and lack of uncertainty modeling.
- The proposed method maintained formation errors (FRE) within acceptable bounds (e.g., $0.10 $–$ 0.22$m) even in GPS-denied environments and under wind disturbances.
Scalability & Computation:
- Computation time per agent remained low and did not scale significantly with the number of agents (tested up to 10 agents in simulation).
- The system ran successfully on single-board computers (Qualcomm Flight Pro) with limited resources.
Robustness:
- The system successfully operated in a metal dome (GPS-denied) and under tree canopies.
- It demonstrated resilience to NLOS conditions (foam/cardboard obstacles) where UWB signals persisted despite visual occlusion.

5. Significance

This work represents a significant step toward fully autonomous, infrastructure-free multi-robot systems. By moving away from centralized architectures and rigid formation control, the authors demonstrate that:

Robustness: Teams can operate in "denied" environments (no GPS, no anchors) where traditional methods fail.
Efficiency: Decentralized, asynchronous processing allows for scalable deployment on low-cost hardware.
Safety: Explicitly modeling uncertainty in the control loop prevents the aggressive, unstable behaviors often seen in formation control when sensors are noisy.

The proposed framework provides a practical solution for applications such as search and rescue, environmental monitoring, and collaborative inspection in complex, unstructured environments.