Input Design for Angles-Only Initial Relative Orbit Determination Using Active Learning

arXiv:2505.21248 · eess.SY, cs.SY · Submitted 2025-05-27 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Input Design for Angles-Only Initial Relative Orbit Determination Using Active Learning".

Dev: Here's the gist of the paper. Accurate relative orbit determination is a significant challenge in modern space operations when relying only on angular measurements, and this work proposes a hybrid estimation and control strategy that uses active learning to design initial input sequences to maximize observability <ref:2505.21248#pg2>.

Rosa: Problem Formulation of Angle-Only IROD The problem of angle-only IROD consists of estimating the relative position r(t0) and relative velocity v(t0) between a chaser and a target spacecraft at the initial observation time t0 <ref:2505.21248#pg5>.

Dev: The chaser, equipped with a passive optical camera, measures only the angular position of the non-cooperative target <ref:2505.21248#pg7>.

Rosa: This measurement provides only angular measurements (azimuth and elevation) of the target, therefore lacking direct range information <ref:2505.21248#pg4>.

Title and authors: Rosa: To get into the main point of "Input Design for Angles-Only Initial Relative Orbit Determination Using Active Learning," they are proposing an active learning input design to generate control sequences offline that optimally balance exploration—finding new information about the initial state—and exploitation, which is keeping things stable <ref:2505.21248#pg4>.

Dev: They frame this whole thing as an optimal experiment design problem, where the learner actively queries the environment by selecting control inputs specifically to get the most informative measurements for parameter or state estimation <ref:2505.21248#pg4>.

Taro: So instead of just guessing what maneuver to do next based on some intuition about the orbit, this method designs a sequence that is mathematically optimized to improve our knowledge about where we are and how fast we’re moving?

Rosa: And they’ve done something quite technical by extending the analytical batch IROD solution from a previous paper to explicitly handle generic impulsive control input sequences <ref:2505.21248#pg4>.

Dev: That extension is crucial because it lets them resolve that scale ambiguity—that problem where you don't know if you are one hundred meters away or one hundred kilometers away—and they derive an analytical covariance for that initial relative state estimate <ref:2505.21248#pg4>.

The paper's summary: Rosa: The main improvement they highlight is this whole hybrid architecture, which splits the process into two distinct phases: first, a phase dedicated to getting that accurate initial state estimate using the Active Learning-Enhanced IROD.

Dev: And then there’s the second phase, which is pure rendezvous control where an Extended Kalman Filter takes over with that batch estimate and feeds it to a Model Predictive Controller <ref:2505.21248#pg4>.

Taro: So it starts with this batch estimate from active learning, and only when the uncertainty gets low enough does the system switch over to real-time recursive filtering and control, which is where things get tricky?

Rosa: That transition point is determined by an analytical covariance metric they derived, which serves both as a quantitative measure of estimation quality and a direct metric for how observable the initial state truly is <ref:2505.21248#pg4>.

Dev: They use singular value decomposition on this normalized covariance to get a condition number, which tells you how much information is distributed across the state space; a lower number means the excitation spreads information more evenly <ref:2505.21248#pg12>.

The paper's improvements: Rosa: To wrap up, this paper introduces this comprehensive strategy for autonomous spacecraft rendezvous using angle-only measurements by combining active learning input design with a batch least-squares IROD algorithm <ref:2505.21248#pg4>.

Dev: They use those active learning inputs to resolve the scale ambiguity and then derive an analytical error covariance to set a clear threshold for when to trust that initial batch estimate enough to start the real-time Extended Kalman Filter and Model Predictive Controller loop <ref:2505.21248#pg4>.

Taro: It’s interesting how they link the batch estimation phase, which is deterministic, with the subsequent recursive filtering phase, which handles all that uncertainty of real-time operation?

Rosa: The performance validation shows that this MPC with active learning strategy consistently outperforms other methods, maintaining a relative mean absolute error below five percent up to six kilometers for initial distances <ref:2505.21248#pg20>.

Dev: And they showed that this method provides a way to design inputs online by using sequential active learning, which lets the system adapt to evolving measurement conditions during the actual rendezvous phase <ref:2505.21248#pg4>.

Taro: It’s good to see a method that doesn't just rely on predefined maneuvers but actively seeks out better measurements when things get tricky out there?

Rosa: The paper, "Input Design for Angles-Only Initial Relative Orbit Determination Using Active Learning," gives us a solid blueprint for tackling the observability limits of angle-only measurements in space robotics <ref:2505.21248#pg1>.

Conclusion: Rosa: So, to wrap up "Input Design for Angles-Only Initial Relative Orbit Determination Using Active Learning," this paper gives us a full strategy for autonomous spacecraft rendezvous using angle-only measurements.

Dev: They use active learning input design to solve that scale ambiguity and then use an analytical error covariance to decide when it’s safe to switch from the batch estimate to the real-time Extended Kalman Filter loop.

Taro: It’s interesting how they link that deterministic batch phase with the subsequent recursive filtering phase, which handles all the uncertainty of real-time operation.

Rosa: The performance validation shows that this Model Predictive Control with active learning strategy consistently beats other methods, keeping the relative mean absolute error below five percent for initial distances up to six kilometers.

Dev: And they showed that this method allows you to design inputs online by using sequential active learning, which lets the system adapt to changing measurement conditions during the actual rendezvous phase.

Taro: It’s good to see a method that doesn't just rely on predefined maneuvers but actively seeks out better measurements when things get tricky out there.

Rosa: That’s what it boils down to, figuring out how to get reliable relative motion estimates when you only have angular data from the ground.

Dev: It shows that even with limited information, if you design your probing inputs smartly, you can bootstrap a whole closed-loop control system reliably.

Taro: The real question is how long this online active learning keeps working before the orbital dynamics become too complex for it to keep up with the required loop rate.

Rosa: That’s definitely where future work needs to focus, figuring out those specific safety thresholds for that transition point we talked about.

Dev: Exactly, because if you can nail that boundary between batch and real-time estimation, you unlock truly robust proximity operations.

Kui Xie, Giovanni Romagnoli, Giordana Bucchioni, Alberto Bemporad

IMT School for Advanced Studies Lucca · University of Pisa

eess.SY, cs.SY

Submitted: 2025-05-27

Updated: 2026-10-27

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 79/100

The gist: Accurate relative orbit determination is a significant challenge in modern space operations when relying only on angular measurements…

Key concepts

Angle-Only IROD
This is the challenge of determining the relative position and velocity between two spacecraft when only angular measurements (azimuth and elevation) are available. Because range information is missing, this problem is inherently difficult because simple measurements don't guarantee a complete picture of their relative motion.
Active Learning (AL) for Input Design
Instead of guessing control maneuvers, AL treats input design like a learning task. The system actively selects the next control input to be applied in a way that yields the most informative measurement possible, efficiently exploring the unknown state space to improve estimation accuracy.
Analytical Covariance and Observability Metric
The paper uses an analytical covariance matrix derived from the IROD solution to measure how well-conditioned the estimation problem is. By analyzing its singular values, it quantifies observability—how effectively the chosen inputs spread information across all possible state uncertainties.
Closed-Loop Rendezvous Architecture
The system operates in two stages: first, an Active Learning phase provides a high-accuracy batch estimate. Once uncertainty drops below safety limits, this solution initializes a real-time Extended Kalman Filter (EKF). The EKF then feeds estimates to a Model Predictive Controller (MPC) for continuous, closed-loop control during the final approach.

Terminology

Summary

Accurate relative orbit determination is a significant challenge in modern space operations when relying only on angular measurements, and this work proposes a hybrid estimation and control strategy that uses active learning to design initial input sequences to maximize observability <ref:2505. p. 021248#pg2>.

Problem Formulation of Angle-Only IROD

The problem of angle-only IROD consists of estimating the relative position r(t0) and relative velocity v(t0) between a chaser and a target spacecraft at the initial observation time t0 <ref:2505.21248#pg5>. The chaser, equipped with a passive optical camera, measures only the angular position of the non-cooperative target <ref:2505. p. 321248#pg7>. This measurement provides only angular measurements (azimuth and elevation) of the target, therefore lacking direct range information <ref:2505. p. 121248#pg4>. This lack of direct range information gives rise to the angle-only initial relative orbit determination (IROD) problem, which is inherently challenging due to the associated observability limitations <ref:2505. p. 121248#pg4>. Specifically, angle-only measurements do not guarantee full observability of the relative chaser-target state in the absence of additional information or excitation <ref:2505. p. 121248#pg4>.

Active Learning for Input Design

An alternative approach relies on exploiting the natural relative orbital dynamics to eventually achieve observability over time, but this method remains inherently ill-conditioned <ref:2505.21248#pg4>. A more principled alternative to heuristic maneuvers is Optimal Experiment Design (or Input Design), which seeks to generate excitation signals that maximize the information content of the measured data for parameter or state estimation <ref:2505. p. 121248#pg4>. In the context of machine learning, this paradigm is referred to as Active Learning (AL), where the learner actively queries the environment, in this case, by selecting control inputs to obtain the most informative measurements <ref:2505. p. 121248#pg4>. The main contributions of this work are: (i) the formulation of the input design problem as an AL task, leveraging concepts from Optimal Experiment Design [10, 11] to compute offline input sequences that optimally trade off exploration (enhancing initial-state observability) and exploitation (enforcing station keeping), within a dual-control framework <ref:2505. p. 121248#pg4>. (ii) the extension of the AL framework to the dynamic setting ([12, 13]), explicitly accounting for the relative orbital dynamics in the design of informative excitation maneuvers <ref:2505. p. 121248#pg4>. (iii) the extension of the analytical batch IROD solution of [15] to explicitly handle generic impulsive control input sequences, enabling the resolution of the scale ambiguity and yielding a unique initial relative state estimate, together with the derivation of its analytical covariance <ref:2505. p. 121248#pg4>.

Analytical Covariance and Observability Metric

The analytical covariance matrix P0, derived in Sec. III.B, serves as this quantitative measure, relating to the Cramér-Rao lower bound, which states that P0 ≥ I−1 (where I is the Fisher Information Matrix) <ref:2505. p. 221248#pg10>. To establish a rigorous basis for the observability metric, we employ Singular Value Decomposition (SVD) on the normalized covariance [19] <ref:2505. p. 221248#pg11>. The degree of observability is determined by the disparity between the most and least uncertain directions, which is captured by the condition number κ(P0,norm) = σmax / σmin = σ1 / σ6 <ref:2505.21248#pg12>. A lower value indicates that the excitation provided by the inputs distributes information more uniformly across the state space, leading to a better-conditioned estimation problem and enhanced robustness <ref:2505. p. 121248#pg4>.

Closed-Loop Rendezvous Architecture

The proposed hybrid architecture separates the problem into two phases: an initial phase dedicated to obtaining an accurate initial state estimate, and a subsequent phase for tracking and control <ref:2505.21248#pg4>. Stage 1 is the Active Learning-Enhanced IROD, which provides a batch estimate of the initial relative state and its analytical covariance <ref:2505. p. 221248#pg4>. Stage 2 is Rendezvous, where an Extended Kalman Filter (EKF) is initialized with the IROD solution and takes over for sequential estimation, providing state estimates to a Model Predictive Controller (MPC) to complete the rendezvous <ref:2505. p. 021248#pg4>. The transition to closed-loop operation is triggered when the maximum eigenvalues of the covariance matrix satisfy safety thresholds <ref:2505. p. 221248#pg14>.

Numerical Validation and Performance

The MPC with AL strategy consistently outperforms the other two strategies, demonstrating robust performance across varying operational conditions <ref:2505.21248#pg20>. The MPC with AL design maintains RMAE below 5% up to 6 km for initial distances and across all tested measurement intervals <ref:2505. p. 221248#pg20>. The simulation validated each component of the pipeline, showing that the AL design consistently improved estimation accuracy over baseline strategies <ref:2505. p. 021248#pg20>. The EKF, initialized by the AL-enhanced IROD solution, enables reliable closed-loop control during proximity operations <ref:2505. p. 221248#pg4>.

Conclusion

The method is built upon a batch least-squares IROD algorithm that exploits known impulsive inputs to resolve the inherent scale ambiguity of angle-only measurements <ref:2505.21248#pg4>. The excitation inputs are obtained with an AL-based input design algorithm that optimizes observability while respecting station-keeping constraints <ref:2505. p. 121248#pg4>. A key contribution lies in the analytical derivation of the estimation error covariance associated to the IROD solution, which provides a direct and rigorous metric to quantify observability <ref:2505. p. 221248#pg4>. The covariance is then used to determine when the batch estimate is accurate enough to bootstrap real-time operations and initialize an EKF together with the propagated IROD solution <ref:2505. p. 221248#pg4>. Finally, the EKF provides sequential state estimates to an MPC, closing the loop for the final rendezvous phase <ref:2505. p. 021248#pg4>. The paper presents a comprehensive strategy for autonomous spacecraft rendezvous using angle-only measurements <ref:2505. p. 321248#pg2>.

Funding Sources

This work was partially supported by the Italian Ministry of Education and Research (MUR) within the framework of the FoReLab project (Departments of Excellence), by the Italian Space Agency (Agenzia Spaziale Italiana, ASI) under the research project DORA - Determinazione Orbitale Relativa Autonoma on-board, con applicazioni terrestri e lunari, through agreement no. 2024-35-HH.0 (CUP n. F53C2400036001) between ASI and the University of Pisa, and by the European Union (ERC Advanced Research Grant COMPACT, No. p. 0 101141351) <ref:2505. p. 021248#pg26>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved initial state estimation via Active Learning-Enhanced IROD: The system can design optimal input control sequences by maximizing exploration of output space to enhance observability, overcoming inherent scale ambiguity in angle-only measurements.

  2. Analytical covariance analysis for robust initialization: The system derives an analytical expression for the estimation error covariance, allowing it to quantify observability and determine when a batch estimate is accurate enough to bootstrap real-time operations.

  3. Transition from batch estimation to sequential filtering: The framework allows the system to transition from a deterministic batch estimate (IROD) to a recursive Extended Kalman Filter (EKF) initialization when uncertainty falls below mission-defined safety thresholds, ensuring reliable closed-loop control.

  4. Constraint-aware Model Predictive Control for rendezvous: The MPC component can compute optimal control actions while explicitly handling state and actuator constraints, optimizing fuel efficiency during the approach phase by minimizing a cost function that balances tracking error and control effort.

  5. Real-time adaptation via sequential active learning: The system can select the next control input online to maximize exploration of the output space based on current state estimates, enabling real-time adaptation to evolving measurement conditions and state uncertainties.

Related papers