Active Learning-Based Input Design for Angle-Only Initial Relative Orbit Determination

Imagine you are trying to meet a friend in a massive, pitch-black foggy field. You can't see them, and they can't see you. The only thing you have is a flashlight that tells you which direction your friend is (left, right, up, down), but it gives you no information about how far away they are.

This is the exact problem spacecraft face when trying to dock with another satellite or clean up space debris using only optical cameras. The camera sees the angle, but not the distance. This is called the "Angle-Only" problem. Without knowing the distance, the spacecraft doesn't know if its friend is 10 meters away or 10 kilometers away. This is a dangerous game of "guess who."

This paper proposes a clever, automated solution to solve this guessing game and safely guide two spacecraft together. Here is how it works, broken down into simple steps:

1. The Problem: The "Scale Ambiguity"

If you just sit still and look at your friend in the fog, you can never figure out the distance. If you move a little bit, you might get a slightly better clue, but if you move in a predictable way, you still might not know the truth.

In space, if a spacecraft just drifts naturally, the math says: "It could be 100 meters away moving slowly, OR it could be 1,000 meters away moving 10 times faster." Both scenarios look exactly the same to the camera. This is called observability. The system is "blind" to the true scale.

2. The Solution: "Active Learning" (The Smart Dancer)

Instead of waiting passively or making random, jerky movements (which wastes fuel), the authors teach the spacecraft to dance.

They use a concept called Active Learning. Think of it like a detective trying to solve a mystery.

Passive Detective: Sits in a chair and waits for clues to walk in. (Bad idea in space; you might never get a good clue).
Active Detective: Gets up, walks around the room, opens drawers, and shakes things to see what falls out. (This is what the spacecraft does).

The spacecraft calculates a specific, pre-planned series of tiny thruster burns (a "dance routine") designed specifically to confuse the ambiguity. By moving in a very specific, non-random pattern, it forces the geometry of the situation to change in a way that makes the distance obvious. It's like holding your thumb up and closing one eye, then the other; the way your thumb "jumps" against the background tells your brain exactly how far away it is.

3. The Two-Stage Strategy

The paper proposes a hybrid system that works in two phases, like a relay race:

Phase 1: The "Batch" Detective (The Initial Guess)

The Goal: Get a rough but accurate fix on where the other ship is.
The Method: The spacecraft executes the "Active Learning" dance routine. It takes a series of photos, runs a complex math calculation (Least Squares), and solves the puzzle.
The Safety Check: The computer calculates a "confidence score" (Covariance). It asks: "Are we sure enough to stop guessing and start driving?" If the uncertainty is too high, it keeps dancing. If the score is good enough, it passes the baton.

Phase 2: The "Real-Time" Driver (The Final Approach)

The Goal: Actually dock with the target.
The Method: Once the initial guess is good enough, the system switches to an Extended Kalman Filter (EKF). Think of this as a super-fast, real-time GPS that updates its position every second based on new photos.
The Steering: This GPS feeds data to a Model Predictive Controller (MPC). The MPC is like an autopilot that looks ahead, plans the most fuel-efficient path to the target, and steers the ship smoothly to a gentle stop.

4. Why This is Better Than Old Methods

Old Way: "Let's just wiggle the ship randomly" (Dithering). This wastes fuel and might not actually solve the distance problem.
Old Way: "Let's just sit still and hope the natural drift of the orbit helps." This takes too long and is unreliable.
This Paper's Way: "Let's calculate the perfect wiggle that solves the math puzzle in the shortest time while using the least amount of fuel."

The Big Picture

The authors tested this in computer simulations. They showed that their "Active Learning" dance routine:

Solved the distance mystery much faster and more accurately than random movements.
Kept the ship safe by not drifting too far away from its intended path.
Successfully guided the ship from a blind guess to a perfect, gentle docking.

In summary: This paper teaches spacecraft how to be smart detectives. Instead of waiting for the fog to clear, they perform a specific, calculated dance that clears the fog for them, allowing them to safely meet up with their target without needing expensive radar or help from Earth. It's about turning a "blind guess" into a "confident handshake" using math and smart movement.

Here is a detailed technical summary of the paper "Active Learning-Based Input Design for Angle-Only Initial Relative Orbit Determination" by Xie et al.

1. Problem Statement

The paper addresses the challenge of Angle-Only Initial Relative Orbit Determination (IROD) for autonomous spacecraft rendezvous.

Context: Modern space operations (docking, debris removal) increasingly rely on passive optical cameras due to their low mass, power, and size. However, these sensors only provide angular measurements (azimuth and elevation), lacking direct range information.
Core Difficulty: Angle-only measurements suffer from an inherent scale ambiguity. Without additional information or active excitation, the relative state (position and velocity) is unobservable because the system dynamics are invariant to a positive scaling of the initial state. This leads to divergent estimates and mission failure risks.
Limitations of Existing Solutions: Current methods often rely on heuristic maneuvers, stereo vision (increasing hardware complexity), or target feature recognition (requiring prior knowledge). These are often suboptimal, fuel-inefficient, or ineffective in specific geometries.

2. Methodology

The authors propose a hybrid estimation and control framework that integrates three distinct stages:

A. Active Learning (AL) for Input Design

To resolve observability issues, the paper formulates the input design problem as an Active Learning task.

Goal: Design an offline sequence of impulsive control inputs ( $\mathbf{u}$ ) that maximizes the information gain regarding the unknown initial state while maintaining station-keeping constraints.
Approach: The authors propose a Dual-Control framework. They optimize an objective function that balances:
1. Exploration: Maximizing the "spread" of measurements in the output space (output space exploration) to break scale ambiguity.
2. Exploitation: Minimizing deviation from a reference trajectory (station-keeping) and control effort.
Algorithms: Two specific AL strategies are compared:
- Expected Error Minimization: Minimizes the expected state estimation error over a set of sampled initial states.
- Greedy- $y$ (Batch AL): Maximizes the dispersion of predicted measurements by minimizing the sum of inverse squared distances between output samples. This avoids the bilevel structure of error minimization, improving computational efficiency.

B. Batch IROD with Analytical Covariance

Once measurements are collected under the designed input sequence:

Estimation: A batch Least-Squares (LS) solution is derived to solve for the scale factors ( $k$ ) that resolve the range ambiguity, yielding a unique initial relative state estimate ( $\hat{\mathbf{x}}_0$ ).
Covariance Derivation: A key theoretical contribution is the derivation of an analytical error covariance matrix ( $\mathbf{P}_0$ ) for the batch LS solution. This accounts for sensor noise propagation through the nonlinear mapping from angles to state.
Observability Metric: The authors use the condition number of the normalized covariance matrix as a quantitative metric for observability. A lower condition number indicates a well-conditioned estimation problem with uniform uncertainty distribution.

C. Transition to Sequential Control

The framework transitions from batch processing to real-time recursive control:

Validation: The batch estimate is accepted only when the maximum eigenvalues of the position and velocity covariance sub-matrices fall below mission-defined safety thresholds.
Handover: The validated batch estimate and its propagated covariance initialize an Extended Kalman Filter (EKF).
Control: The EKF provides real-time state estimates to a Model Predictive Controller (MPC), which handles the final approach, explicitly managing state and actuator constraints to ensure a safe rendezvous.

3. Key Contributions

Active Learning Formulation: The first application of Active Learning concepts to the angle-only IROD problem, treating input design as an optimization task to maximize observability dynamically.
Analytical Covariance for Batch IROD: Derivation of a closed-form analytical covariance matrix for the batch LS solution under impulsive inputs. This provides a rigorous, quantitative metric to assess estimation quality and determine the optimal transition point to recursive filtering.
Hybrid Architecture: A complete end-to-end pipeline (AL Design $\to$ Batch IROD $\to$ EKF $\to$ MPC) that bridges the gap between initial coarse estimation and precise terminal control.
Robustness to Model Mismatch: The system is validated using nonlinear Keplerian dynamics for the "true" spacecraft motion while the estimator uses linear Clohessy-Wiltshire (CW) equations, demonstrating robustness to model errors.

4. Results

Numerical simulations were conducted in a Low Earth Orbit (LEO) scenario with a V-bar station-keeping configuration (a worst-case scenario for observability).

Estimation Accuracy: The MPC with AL strategy significantly outperformed baselines (MPC only and MPC with random dithering). It achieved a Relative Mean Absolute Error (RMAE) of < 5% for initial distances up to 6 km, whereas baselines failed to resolve scale ambiguity or exhibited large deviations.
Covariance Validation: The analytical covariance bounds derived in the paper accurately bounded the empirical estimation errors in Monte Carlo simulations (100 runs), validating the theoretical derivation.
Conditioning: The AL strategy produced the lowest condition numbers for the covariance matrix, indicating superior numerical stability and information content compared to heuristic methods.
Closed-Loop Performance: The hybrid system successfully executed a full rendezvous. The EKF, initialized by the AL-enhanced IROD, converged rapidly to true states, and the MPC guided the chaser to the target with final errors of $\Delta p \approx 9.87$ mm and $\Delta v \approx 5.37$ mm/s.

5. Significance

This work provides a critical step toward fully autonomous, resource-constrained space operations.

Autonomy: It enables small satellites to perform complex rendezvous and proximity operations without relying on external aids (like GPS or cooperative transponders) or heavy hardware (stereo cameras).
Safety: By rigorously quantifying observability via analytical covariance, the system ensures that the transition to closed-loop control only occurs when the state estimate is sufficiently accurate, mitigating collision risks.
Efficiency: The Active Learning approach optimizes fuel usage by generating maneuvers that are specifically tailored to maximize information gain, rather than using arbitrary or suboptimal excitation patterns.

In summary, the paper presents a mathematically rigorous and practically validated framework that solves the fundamental scale ambiguity problem in angle-only navigation, enabling reliable autonomous rendezvous for next-generation space missions.