💻 computer science

Attitude Estimation Using Inertial and Barometric Measurements

This paper proposes two barometer-aided attitude estimation architectures for autonomous vehicles operating in GNSS-denied or highly accelerated conditions, utilizing nonlinear observers on SO(3) and SO(3)×R² to achieve stable and robust tilt estimation by resolving the ambiguity between gravitational and inertial accelerations.

Original authors: Melone Nyoba Tchonkeu, Soulaimane Berkane, Tarek Hamel

Published 2026-07-16

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Melone Nyoba Tchonkeu, Soulaimane Berkane, Tarek Hamel

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 trying to teach a robot how to fly without ever looking at the sky. You give it a tiny compass to find North, a barometer to feel changes in air pressure, and a gyroscope to feel how it's spinning. But here's the tricky part: if the robot suddenly speeds up, turns sharply, or gets hit by a gust of wind, its internal motion sensors get confused. It can't tell the difference between the force of gravity pulling it down and the force of its own engine pushing it forward. To a robot, a sudden dive feels exactly like a sudden turn. This is the "tilt problem." If the robot thinks it's tilting when it's actually just speeding up, it will crash. Usually, engineers fix this by adding a GPS or a fancy radar to tell the robot how fast it's moving through the air. But what if those tools are broken, too expensive, or simply not there?

This is where the story gets interesting. Scientists have long wondered if there is a simpler, cheaper way to solve this puzzle. Enter the barometer. You might know it as the thing in your phone that tells you which floor of a building you are on. It measures air pressure, which changes as you go up or down. While a barometer can't tell you how fast you are moving forward, it is incredibly good at telling you exactly how you are moving up and down relative to the ground. This paper explores a clever idea: what if we use that simple up-and-down information to help the robot figure out which way is "down" (gravity), even when it's doing crazy, fast maneuvers? By combining the robot's internal motion sensors with this humble pressure sensor, the researchers built a new kind of "brain" that helps drones and robots stay upright without needing expensive GPS or wind sensors.

The Problem: When "Up" Becomes "Forward"

For a flying robot, knowing its attitude—which way is up, down, left, or right—is the most important job it has. If it gets this wrong, it crashes. The robot usually relies on an IMU (Inertial Measurement Unit), a tiny chip that acts like an inner ear. It has a gyroscope to feel spinning and an accelerometer to feel pushes and pulls.

The problem is that the accelerometer is a bit of a liar. It measures "specific acceleration," which is a fancy way of saying "all the forces acting on me except gravity." When a robot is sitting still, the accelerometer feels the Earth pulling it down, so it knows which way is "down." But when the robot accelerates forward, the accelerometer feels that push too. It gets confused. It thinks, "Am I tilting forward, or am I just speeding up?" Without a clear answer, the robot loses its balance.

Usually, engineers fix this by adding a velocity sensor (like GPS or a radar tube) to tell the robot exactly how fast it is moving. But these sensors can be heavy, expensive, or fail in bad weather. The authors of this paper asked: Can we solve the tilt problem using only a barometer?

The Solution: Two New Ways to Think

The researchers designed two different "observers" (mathematical brains) to solve this puzzle. Think of these observers as two different strategies for a detective trying to solve a crime.

Strategy 1: The Two-Step Detective (AGAS)
The first approach is like a detective who works in two distinct phases.

Phase One: A specialized tool (a Riccati observer) looks at the barometer and the IMU to figure out just two things: how high the robot is and how fast it is moving up or down. It also guesses the direction of "down" (tilt).
Phase Two: Once that guess is made, a second tool (a nonlinear observer on SO(3)) takes that guess and combines it with a magnetometer (compass) to figure out the robot's full 3D orientation.

The authors proved mathematically that this method is incredibly robust. Even if the robot starts with a huge mistake in its head (like thinking it's upside down when it's actually right side up), this two-step detective will eventually figure out the truth. They call this Almost-Global Asymptotic Stability (AGAS). It's like saying, "No matter where you start in the world, you will eventually find your way home, even if you have to walk a long, slow path to get there."

Strategy 2: The Fast-Track Detective (LES)
The second approach is a unified, all-in-one detective. Instead of working in steps, it fuses the barometer, the compass, and the IMU all at once into a single mathematical model.
This method is much faster. If the robot starts with a small mistake, this detective zooms to the correct answer exponentially fast. The authors call this Local Exponential Stability (LES). However, there is a catch: this detective only works if the robot is moving in a specific way. It needs the robot to wiggle or move horizontally enough to "shake" the sensors and reveal the truth. If the robot is just hovering perfectly still, this fast-track detective might get stuck.

The Experiment: Simulations and Real Flights

To see if these ideas actually work, the team didn't just write equations; they tested them.

First, they ran simulations on a computer. They created a virtual robot and threw it into two scenarios:

Small Mistakes: The robot started with a tiny error. Here, the Fast-Track (LES) detective won hands down. It corrected the error in about 9 seconds, while the Two-Step (AGAS) detective took about 14 seconds.
Huge Mistakes: The robot started with a massive error (like a 60-degree tilt). Both detectives eventually fixed it, but the Two-Step detective was smoother and more reliable, while the Fast-Track detective had a bit of a wobble before it settled down.

Then, they took it to the real world. They flew a real drone (a MakeFlyEasy Fighter VTOL) equipped with these sensors. They intentionally gave the robot a terrible starting guess (a 60-degree error) to see if it could recover.

The Result: Both methods worked. The robot recovered its balance and flew straight.
The Trade-off: The Fast-Track (LES) observer was faster to correct itself and had smaller errors during the recovery phase. However, the Two-Step (AGAS) observer was slightly more accurate in the long run once everything was settled.

What This Means

The paper suggests that you don't always need expensive GPS or wind sensors to keep a drone stable. A simple, cheap barometer can do the heavy lifting if you use the right math.

The authors found that there is a trade-off:

If you need speed and your robot is moving around a lot, the LES (Fast-Track) method is superior. It corrects errors quickly, provided the robot is moving enough to provide "excitation" (movement data).
If you need reliability and your robot might start with a huge error or move very slowly, the AGAS (Two-Step) method is safer. It guarantees the robot will eventually find the truth, even if it takes a little longer.

In the end, this research offers a lightweight, practical toolkit for engineers. It shows that by listening carefully to the air pressure, a robot can know which way is up, even when the wind is blowing and the engines are roaring.

Technical Summary: Attitude Estimation Using Inertial and Barometric Measurements

Problem Statement
Accurate attitude estimation for autonomous vehicles remains a critical challenge, particularly in Global Navigation Satellite System (GNSS)-denied environments or during highly accelerated flight. In such conditions, Inertial Measurement Units (IMUs) alone are insufficient for reliable tilt estimation because accelerometers measure specific acceleration (the sum of non-gravitational forces and gravity), creating an ambiguity between gravitational and inertial accelerations. While velocity-aided attitude (VAA) approaches using GNSS, Pitot tubes, or Doppler radar are common, these sensors may be unavailable, intermittent, or sensitive to airflow disturbances. The paper addresses the need for a lightweight, robust alternative that exploits barometric altitude measurements to enhance attitude estimation without relying on full velocity vectors.

Methodology
The authors propose two complementary barometer-aided attitude estimation architectures, both operating within geometric observer frameworks on the Special Orthogonal Group $SO(3)$ and its extensions. The vehicle dynamics are modeled using standard translational and attitude kinematics, with the barometer providing altitude measurements ( $h$ ) and a magnetometer providing Earth's magnetic field references.

Two-Stage Observer (AGAS):
- Architecture: This design cascades a deterministic Riccati observer with a nonlinear observer on $SO(3)$.
- Mechanism: The first stage estimates the vertical motion (altitude and vertical velocity) and the tilt (gravity direction in the body frame) using IMU and barometric data. The second stage fuses this tilt estimate with magnetometer measurements to recover the full attitude.
- Stability: Under a Uniform Observability (UO) condition requiring the full inertial acceleration vector to be persistently exciting, this architecture guarantees Almost-Global Asymptotic Stability (AGAS). It preserves the geometric structure of the attitude dynamics.
One-Stage Observer (LES):
- Architecture: This is a unified reduced-order nonlinear observer evolving on $SO(3) \times \mathbb{R}^2$ .
- Mechanism: It directly integrates IMU, barometric, and magnetometer measurements into a single framework. The innovation terms are generated by a local Riccati equation derived from linearized error dynamics.
- Stability: This architecture guarantees Local Exponential Stability (LES). Crucially, its UO condition is relaxed: it only requires persistent excitation of the horizontal component of the inertial acceleration, rather than the full vector. This makes it less restrictive regarding vehicle motion.

Both designs include discrete-time implementations suitable for real-time execution on embedded systems, utilizing exponential-Euler integration for the attitude components.

Key Contributions

Novel Sensing Modality: The paper establishes that barometric altitude measurements provide sufficient complementary information to constrain the gravity direction for tilt estimation, even in the absence of conventional velocity measurements.
Dual Observer Designs: The authors present two distinct architectures:
- A two-stage observer ensuring AGAS, suitable for scenarios where global convergence from arbitrary initial conditions is prioritized.
- A unified one-stage observer ensuring LES, which operates under significantly weaker observability conditions (requiring only horizontal excitation).
Theoretical Analysis: The work provides a rigorous characterization of the observability properties for barometer-aided estimation, establishing the trade-off between the domain of attraction (AGAS) and the convergence rate/observability requirements (LES).
Validation: Theoretical findings are corroborated by extensive Monte Carlo simulations and real-flight experiments using a VTOL UAV.

Results

Simulations: In scenarios with small initial errors, the LES observer demonstrated significantly faster convergence (approx. 9s vs. 14s) and a smaller transient envelope compared to the AGAS observer. In scenarios with large initial errors, both observers converged, though the AGAS observer exhibited a smoother transient with reduced overshoot, while the LES observer showed larger initial oscillations before entering its region of attraction.
Experimental Flight Data: Using data from a MakeFlyEasy Fighter VTOL UAV, both observers successfully recovered attitude from large initialization errors (approx. 60° perturbations).
- The LES observer achieved faster convergence times and lower peak attitude errors during the transient phase.
- The AGAS observer demonstrated robust recovery from large errors and achieved slightly better steady-state accuracy in roll and yaw.
- Both observers maintained bounded errors and converged to the ground truth, validating their effectiveness in GNSS-degraded environments.
Observability Analysis: The experiments confirmed that the LES observer's relaxed observability condition allows it to perform effectively even when the excitation matrix is ill-conditioned (e.g., during weak maneuvers), whereas the AGAS observer relies on richer motion for optimal performance.

Significance and Claims
The paper claims to provide a practical, lightweight alternative to conventional velocity-aided attitude estimation for autonomous vehicles operating in environments where velocity sensors are unavailable or degraded. By exploiting barometric data, the proposed methods offer a viable solution for tilt estimation without the cost or complexity of additional velocity sensors.

The authors emphasize the trade-off between the two approaches: the AGAS design offers strong robustness to large initialization errors (almost-global convergence), while the LES design offers faster convergence and operates under less restrictive motion requirements (relaxed observability). The work represents the first unified treatment of barometer-aided attitude estimation that encompasses observability analysis, observer design with both AGAS and LES guarantees, and validation on real-flight data. The results suggest that the choice between architectures depends on the specific application requirements regarding initialization uncertainty and available vehicle excitation.

The Problem: When "Up" Becomes "Forward"

The Solution: Two New Ways to Think

The Experiment: Simulations and Real Flights

What This Means

More like this