💻 computer science

Estimation of Spacecraft Inertia Tensor Using Attitude-Only Data from Torque-Free Motion

This paper presents a computationally efficient, attitude-only framework for estimating a spacecraft's inertia tensor from torque-free motion that utilizes a Karush-Kuhn-Tucker initialization and exact nonlinear refinement to significantly outperform Extended Kalman Filters in both accuracy and speed across single and multi-arc observation scenarios.

Original authors: Daigo Kobayashi, Vakhtang Putkaradze

Published 2026-08-11

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Daigo Kobayashi, Vakhtang Putkaradze

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 watching a spinning top in a dark room. You can't see the top's insides, you can't touch it, and you can't feel how heavy it is. All you have is a camera that takes pictures of the top's orientation as it twirls. Now, imagine that top is a massive, tumbling satellite floating in space, and you need to know exactly how its weight is distributed inside to predict where it will go next. This is the challenge of "inertia," a concept that describes how an object resists changes to its spin. If you know a satellite's inertia, you can predict its future path, plan how to dock with it, or even figure out if it's broken. Usually, scientists need special sensors called gyroscopes to measure how fast the satellite is spinning to figure this out. But what if those sensors are broken, missing, or if you are looking at a piece of space junk that doesn't have any sensors at all? Can you still figure out its secrets just by watching it spin?

This paper by Daigo Kobayashi and Vakhtang Putkaradze says "yes." They developed a clever new way to guess a spacecraft's internal weight distribution using only a history of its angles—like a movie of its orientation—without needing to know how fast it's spinning or if anyone is pushing it. Think of it like trying to guess the shape of a mystery object inside a sealed box just by watching how the box wobbles when you shake it. The authors created a mathematical "detective" that watches the wobble, uses a precise map of how spinning objects behave (based on centuries-old physics), and solves a puzzle to reveal the hidden weight map. They tested this by simulating a camera watching a tumbling satellite and found that their method is not only incredibly accurate but also lightning-fast compared to the old, clunky ways of doing it.

The Mystery of the Wobbly Satellite

Spacecraft are like giant, complex tops. When they spin in the vacuum of space without any engines firing (a state called "torque-free motion"), they follow strict rules of physics. If you know how a top is shaped and where its weight is concentrated, you can predict exactly how it will spin. But the problem is often the reverse: we see the spin, but we don't know the shape or the weight distribution. This is a huge problem for space agencies. If a satellite breaks or runs out of fuel, it might start tumbling wildly. To catch it, fix it, or move it out of the way, engineers need to know its "inertia tensor." That's just a fancy math term for a map of how the satellite's mass is spread out. If you get this map wrong, your prediction of where the satellite will be in an hour could be completely off, leading to a failed mission or a dangerous collision.

Traditionally, to get this map, engineers rely on gyroscopes—sensors that measure how fast the satellite is spinning. But gyroscopes can break, they can be heavy, and they are useless if you are looking at a piece of space debris that has no sensors at all. Some methods try to guess the spin rate by looking at how the satellite's angle changes over time, but that's like trying to measure the speed of a car by looking at a blurry photo of its position; the math gets messy and full of errors.

The New "Attitude-Only" Detective

The authors of this paper asked a bold question: Can we figure out the inertia tensor using only the angles (attitude) of the spacecraft, without ever knowing the spin speed? Their answer is a resounding yes, provided the spacecraft is spinning freely.

They built a two-step system that acts like a super-smart guesser. First, they use a quick, linear math trick (called a Karush–Kuhn–Tucker or KKT formulation) to get a rough guess of the inertia. It's like looking at the wobble of a spinning top and making a quick estimate of its weight distribution. This guess is fast but not perfect.

Then, the real magic happens. They take that rough guess and refine it using a "nonlinear shooting" method. This is where they use the exact, ancient formulas that describe how a spinning top moves (Euler's equations) but solve them using special mathematical functions called "Jacobi-elliptic functions." Think of these functions as a perfect, unbreakable map of how a spinning object must behave. They also use a technique called the "Magnus expansion" to translate that spinning motion into the language of angles (quaternions) that cameras use.

The system works like this:

The Guess: It starts with a rough idea of the inertia.
The Simulation: It uses the perfect math map to predict how the satellite should spin if that guess were true.
The Comparison: It compares this perfect prediction to the actual "movie" of angles we have.
The Correction: If the prediction doesn't match the movie, it tweaks the guess and tries again, over and over, until the math perfectly matches the observation.

The Results: Fast, Accurate, and Camera-Ready

The authors tested their method in two main ways. First, they simulated a satellite spinning for 500 seconds with a perfect camera that had a tiny bit of noise (like a slightly shaky hand). They compared their new method to the standard "Extended Kalman Filter" (EKF), which is the current gold standard for these problems.

The results were striking. Their new method was about 100 times faster than the EKF. While the EKF took over 11 seconds to crunch the numbers on a laptop, their method did it in less than 0.2 seconds. In terms of accuracy, their method reduced the error by about one order of magnitude (ten times better) compared to the EKF. In their simulations, the error in the inertia map was incredibly small, often less than one-millionth of the total value.

But the real test was making it work with a real-world camera. The authors simulated a "chaser" spacecraft taking photos of a tumbling "target" spacecraft (modeled after the JASON-1 satellite). They used a computer vision system to find specific points on the target (like the corners of solar panels) and calculate its angle from the photos. This introduced real-world errors: shadows, weird lighting, and blurry images.

Even with these messy, real-world camera errors, the method worked. When they used the camera-derived angles instead of perfect data, the error in the inertia map went up, but it was still very usable. Most impressively, they used their estimated inertia map to predict where the satellite would be 10 hours later. The prediction was accurate to within a few degrees, proving that even with a "noisy" camera, they could see far into the future.

Why It Matters

This paper suggests that we don't need expensive, fragile gyroscopes to understand how a spacecraft spins. If we can just take pictures of it, we can figure out its internal weight distribution. This is a game-changer for "non-cooperative" targets—spacecraft that are broken, abandoned, or space debris. It means a future mission could fly up to a piece of junk, snap some photos, calculate exactly how it will tumble, and then safely grab or push it away.

The authors also found some interesting limits. If a satellite is spinning perfectly around one of its main axes (like a perfectly balanced top), it's hard to figure out its weight distribution because it doesn't wobble enough. But if it's tumbling a bit more chaotically, the method shines. They also found that spinning too fast or too slow can make it harder to get a good picture, but for most realistic tumbling scenarios, their "attitude-only" detective is a powerful new tool for space safety.

In short, this paper shows that with the right math and a good camera, we can solve the mystery of a spinning satellite's weight just by watching it dance in the dark.

Technical Summary: Estimation of Spacecraft Inertia Tensor Using Attitude-Only Data from Torque-Free Motion

1. Problem Statement

The accurate determination of a spacecraft's inertia tensor is critical for attitude prediction, maneuver design, and control system implementation. It is also essential for monitoring structural changes (e.g., fuel depletion) and assessing the rotational state of non-cooperative targets during rendezvous, docking, or debris removal.

Existing methods for inertia identification typically rely on gyroscope measurements, known control torques, or active excitation maneuvers. However, these requirements are often prohibitive for:

Non-cooperative space-domain-awareness scenarios: Where direct rate telemetry is unavailable.
Small spacecraft: Equipped with low-cost MEMS gyroscopes that may be unreliable.
Passive observation: Where active excitation is not feasible.

This paper addresses the specific challenge of estimating a spacecraft's normalized inertia tensor using attitude histories alone during torque-free rotational motion, eliminating the need for gyroscope data, known torques, or active maneuvers.

2. Methodology

The proposed framework utilizes a two-tier mathematical approach to propagate the spacecraft's state and an optimization strategy to estimate the inertia tensor.

A. Analytical Phase-Space Mapping

To avoid the truncation errors of discrete numerical solvers (like Runge-Kutta) and the noise sensitivity of differentiating noisy quaternion data, the method employs exact analytical solutions:

Angular Velocity Propagation: The body-frame angular velocity $\boldsymbol{\Omega}(t)$ is propagated using the exact, closed-form solution of Euler's equations for a torque-free rigid body, expressed via Jacobi elliptic functions. This maps the angular velocity from time $t$ to $t+h$ without numerical integration.
Attitude Kinematics: Since a closed-form solution for the attitude quaternion of an asymmetric body does not exist, the method applies the Magnus expansion. By assuming a linear variation of the angular velocity matrix over the time step, the transition matrix for the quaternion is resolved analytically. A second-order truncation ( $m=2$ ) is used, which provides a symplectic and unconditionally stable integrator that preserves the unit-norm constraint of quaternions.

B. Estimation Algorithm

The estimation process is formulated as a nonlinear optimization problem minimizing the mismatch between measured attitudes and those predicted by the analytical model.

KKT Initialization: A fast, linear initialization is performed using a Karush–Kuhn–Tucker (KKT) formulation. This solves a constrained least-squares problem based on the linearized torque-free Euler equations to provide a "warm start" estimate of the normalized inertia tensor (enforcing $\text{tr}(\mathbf{I}) = 1$ ).
Nonlinear Shooting Refinement: The KKT estimate is refined using a nonlinear shooting method. The objective function minimizes the norm-squared of the vector part of the error quaternion between the simulated trajectory (generated via the exact elliptic map and Magnus expansion) and the measured telemetry. This refinement solves for the initial kinematic states and the inertia tensor simultaneously.

C. Image-Based Evaluation

To evaluate performance under realistic conditions relevant to proximity operations, the method was tested using monocular image-derived attitudes.

Simulation: Photorealistic images of a JASON-1 spacecraft were rendered using Mitsuba 3 under various lighting and viewing conditions.
Pose Estimation: A vision pipeline (YOLO detection + RTMPose keypoint detection + EPnP with RANSAC) estimated the target's attitude from the images.
Noise Characteristics: The resulting attitude errors were non-Gaussian, dependent on geometry and illumination, and included outliers handled by reprojection error filtering.

3. Key Contributions

Attitude-Only Framework: A novel method that estimates the normalized inertia tensor without gyroscope data or known torques, relying solely on passive attitude histories.
Exact Analytical Propagation: The integration of exact Jacobi-elliptic solutions for angular velocity with a Magnus-expansion-based quaternion map, reducing sensitivity to measurement noise compared to numerical differentiation.
Efficient Optimization: A hybrid approach combining a fast linear KKT initialization with a precise nonlinear shooting refinement, avoiding the iterative matrix-propagation overhead of Extended Kalman Filters (EKF).
Observability Analysis: A detailed quantification of how rotational excitation affects estimation accuracy, identifying that motions near principal axes (especially stable ones) lead to degraded observability.

4. Results

A. Synthetic Data Performance (Controlled Noise)

Single Arc (500s): The proposed KKT+Elliptic method reduced the mean absolute error (MAE) of the inertia tensor by approximately one order of magnitude compared to an EKF initialized from the same KKT estimate ( $3.56 \times 10^{-7}$ vs. $2.27 \times 10^{-6}$ ).
Computation Time: The proposed method was nearly two orders of magnitude faster than the EKF (0.18s vs. 11.1s).
Multi-Arc (Three 100s arcs): Joint estimation from multiple arcs provided similar accuracy improvements over EKF while remaining significantly faster.
Long-Horizon Prediction: Using the estimated tensor from a 2,000-second single arc, the method supported 10-hour attitude predictions with a median error of single-digit degrees.

B. Image-Derived Data Performance (RPO Scenarios)

Single-Arc RPO: Replacing true attitudes with image-derived estimates increased the median inertia-tensor MAE from $2.0 \times 10^{-4}$ to $6.5 \times 10^{-4}$ . Despite this, the estimated tensor supported 10-hour predictions with a median attitude error of $6.7^\circ$ .
Multi-Arc RPO: Performance was governed by rotational excitation and temporal sampling.
- Low Spin: Provided insufficient excitation, leading to higher errors.
- High Spin: Reduced effective temporal resolution due to fixed image cadence (1s), also degrading performance.
- Medium Spin: Offered the best balance, where the KKT+Elliptic method consistently outperformed both the raw KKT initialization and the KKT+EKF refinement across various arc lengths (30–300 frames).

C. Observability Findings

Principal Axes: Pure rotation about a principal axis yields no information about the tensor structure. Near stable axes, estimation error degrades algebraically ( $O(\epsilon^{-1})$ ) as the motion approaches the axis. Near the unstable intermediate axis, degradation follows a weaker logarithmic trend ( $O(|\log \epsilon|)$ ).
Separatrices: Trajectories near the separatrix (unstable intermediate axis) slow down, reducing the number of observable phase-space traversals within a fixed time window.

5. Significance and Claims

The paper claims that accurate inertia estimation is feasible using only passive optical attitude histories, provided the rotational motion offers sufficient directional excitation. The proposed KKT+Elliptic framework offers a computationally efficient alternative to EKF-based methods, making it suitable for resource-constrained onboard processing.

By connecting torque-free mass-property identification with image-based pose estimation, this work demonstrates a pathway for autonomous inspection, debris characterization, and operations involving tumbling or inactive spacecraft without requiring active excitation or high-fidelity rate sensors. The method is particularly valuable for non-cooperative scenarios where traditional telemetry is unavailable.