🤖 AI

Adaptive Relative Pose Estimation Framework with Dual Noise Tuning for Safe Approaching Maneuvers

This paper proposes an adaptive relative pose estimation framework for Active Debris Removal missions that integrates a CNN-based marker detection system with a dual-noise-tuned Unscented Kalman Filter to robustly estimate the state of tumbling targets like ENVISAT under varying measurement uncertainties and unmodeled dynamics.

Original authors: Batu Candan, Murat Berke Oktay, Simone Servadio

Published 2026-03-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Batu Candan, Murat Berke Oktay, Simone Servadio

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 catch a floating, broken satellite in space with your own spaceship. This is a bit like trying to grab a spinning, tumbling piece of trash floating in a dark, windy room. The problem? The trash (the target satellite) doesn't have a handle, it doesn't talk to you, and the lights keep flickering on and off.

This paper is about building a super-smart "brain" for your spaceship that can figure out exactly where that trash is and how it's spinning, even when things get messy.

Here is the story of how they did it, broken down into simple parts:

1. The Problem: The "Blindfolded" Catch

Usually, to catch something, you need to see it clearly. But in space, the "trash" (like the ESA's ENVISAT satellite) is huge, weirdly shaped, and often gets covered by shadows (eclipses) or other parts of itself.

The Old Way: Traditional navigation systems are like a driver who assumes the road is always straight and the weather is always perfect. If a sudden storm hits or a pothole appears, the driver panics or crashes because they didn't expect it.
The New Way: The authors built a system that acts like a smart, adaptive driver. This driver knows that the road might get bumpy, the fog might roll in, and the car's sensors might glitch. Instead of panicking, the driver adjusts their driving style in real-time.

2. The Eyes: The "AI Detective"

First, the spaceship needs to see the trash. They used a Convolutional Neural Network (CNN), which is basically an AI detective trained to look at photos and find specific corners on the satellite.

The Analogy: Imagine playing "Where's Waldo?" but the picture is blurry, shaking, and half of it is in the dark. The AI detective is really good at spotting the corners of the satellite even when the picture is messy.
The Catch: Even a good detective makes mistakes. Sometimes the AI thinks it sees a corner where there isn't one, or it's a little off because of camera shake.

3. The Brain: The "Smart Filter" (UKF)

Once the AI finds the corners, the spaceship needs to calculate exactly where the satellite is. They used a math tool called an Unscented Kalman Filter (UKF).

The Analogy: Think of the filter as a skeptical coach. The coach listens to the AI detective (the sensor), but the coach also knows how the satellite should be moving based on physics.
- If the detective says, "It's over there!" but the physics say, "No, it's impossible to be that fast," the coach has to decide: Is the detective lying, or is the physics model wrong?
- Traditional coaches just guess a fixed level of trust. If they trust the detective too much, they get confused by bad data. If they trust them too little, they ignore good data.

4. The Secret Sauce: "Dual Noise Tuning"

This is the main breakthrough of the paper. The authors realized that the "trust level" shouldn't be fixed; it should change like a thermostat. They created a Dual-Adaptation Strategy:

Tuning the "Eye" (Measurement Noise):
- Scenario: The camera is shaky, or the satellite is half-hidden. The AI detective is likely to make mistakes.
- The Fix: The system notices the detective is acting jittery. It says, "Okay, I'm not going to trust these numbers as much right now." It lowers its confidence in the sensor data so it doesn't get fooled by the noise.
- Metaphor: It's like when you are trying to hear someone in a loud concert. You lean in and say, "I'm not sure what you said, let me listen again," instead of shouting back a wrong guess.
Tuning the "Physics" (Process Noise):
- Scenario: The satellite goes into a shadow (eclipse) and the camera goes blind. The detective stops talking.
- The Fix: The system knows it can't see anything, so it says, "Okay, I have to guess where the satellite is based on its momentum, but I'm going to be very humble about my guess." It widens its "safety net" (uncertainty) to account for the fact that it might be wrong.
- Metaphor: It's like driving in thick fog. You can't see the road, so you slow down and assume you might be off the road by a few feet. You don't assume you are perfectly centered.

5. The Extra Tool: The "Laser Tape Measure" (LiDAR)

To make things even better, they added a LiDAR sensor (a laser that measures distance).

The Problem: Sometimes the laser hits a spot slightly different from where the AI thinks the corner is. It's like measuring a wall with a tape measure, but the tape is slightly bent.
The Fix: The system adds a "bias" variable. It essentially says, "My laser tape measure is always off by about 2 inches. I will subtract that 2 inches from every measurement." This keeps the math honest.

6. The Results: The "Super Catch"

They tested this system in a super-realistic computer simulation (using a 3D model of the ENVISAT satellite).

Without the new system: The spaceship would get confused during shadows or bad lighting, thinking it knew the position perfectly when it was actually way off.
With the new system: The spaceship stayed calm. When the lights went out, it widened its safety net. When the camera got noisy, it trusted the physics more. When the laser was slightly off, it corrected itself.
The Outcome: The system was much more accurate and stable, even when the satellite was spinning wildly or hidden in darkness.

Summary

This paper is about teaching a spaceship's computer to be flexible. Instead of blindly trusting its eyes or its physics calculations, it constantly asks, "How much should I trust this right now?" and adjusts its confidence accordingly. This makes it much safer and more likely to successfully grab that floating space trash without crashing into it.

1. Problem Statement

The paper addresses the critical challenge of Active Debris Removal (ADR) and on-orbit servicing, specifically focusing on estimating the relative pose (position, velocity, attitude, and angular velocity) of uncooperative targets like the derelict ESA ENVISAT satellite.

Challenges: These missions face severe visibility issues (self-occlusion, eclipses), unpredictable target dynamics, and uncertain sensor performance.
Limitations of Current Methods: Traditional estimation pipelines often rely on fixed noise models (static covariance matrices) within Kalman filters. These fail to adapt to real-time variations in sensor quality (e.g., CNN detection jitter, lighting changes) or unmodeled dynamics (e.g., during eclipse periods when visual data is lost), leading to either over-conservative or over-confident state estimates.

2. Methodology

The authors propose a complete pipeline integrating Computer Vision (CV) with an Adaptive Unscented Kalman Filter (UKF).

A. Simulation and Data Generation

Environment: A high-fidelity simulation was built using Blender (unlike previous MATLAB-based approaches) to generate realistic synthetic images of the ENVISAT satellite.
Conditions: The simulation includes realistic solar illumination, shadowing, and occlusion checks via ray-casting to determine marker visibility.
Ground Truth: The system tracks a 12-component relative state vector (position, velocity, Modified Rodrigues Parameters for attitude, and angular velocity) alongside 2D pixel coordinates and depth values for satellite corners.

B. Vision Processing (CNN)

Model: A lightweight CNN named TinyCornerNET (based on a U-Net++ encoder-decoder with a ResNet-34 backbone) is used to detect structural corners on the target.
Training: Trained on 6,035 synthetic images using focal loss to handle class imbalance between corners and background.
Output: The CNN provides 2D corner coordinates, which are converted to 3D measurements using camera intrinsic/extrinsic parameters and depth information.

C. Filtering Framework (UKF)

State Estimation: The UKF is chosen over the Extended Kalman Filter (EKF) for its ability to handle nonlinear relative dynamics without linearization.
State Vector: 12 states (3 pos, 3 vel, 3 MRP, 3 ang. vel).
Measurement Model: Fuses 2D corner detections (converted to 3D) and, in later stages, LiDAR depth data.

D. The Core Innovation: Dual Adaptive Noise Tuning

The paper introduces a novel dual-adaptation strategy that tunes both the Process Noise Covariance ( $Q$ ) and Measurement Noise Covariance ( $R$ ) online:

Adaptive $R$ Tuning (Measurement Noise):
- Mechanism: Uses an Innovation-based Multiple Tuning Factor (MTF).
- Logic: Monitors the measurement innovation ( $e_k = z_k - \hat{z}_k$ ). If the innovation covariance exceeds the theoretical expectation (indicating high noise or outliers), the diagonal elements of $R$ are increased dynamically.
- Benefit: Prevents the filter from trusting noisy or spurious CNN detections, maintaining robustness during lighting changes or partial occlusions.
Adaptive $Q$ Tuning (Process Noise):
- Mechanism: Inspired by the Rauch–Tung–Striebel (RTS) smoother, it utilizes a forward-looking cross-covariance matrix ( $D_k$ ) derived from sigma points.
- Logic: During periods of no measurement (e.g., eclipse) or model mismatch, the algorithm calculates the divergence between predicted and propagated sigma points. It injects noise into $Q$ specifically in the directions where uncertainty is growing.
- Benefit: Replaces manual, heuristic-based $Q$ inflation during eclipses with a data-driven, real-time adjustment that prevents filter divergence without requiring prior knowledge of eclipse duration.

E. LiDAR Augmentation

To address systematic offsets between projected corners and LiDAR point clouds, the authors augment the state vector with a LiDAR depth bias variable.
A safety gate filters LiDAR data, and the bias is estimated online, significantly improving rotational state accuracy.

3. Key Contributions

Dual-Adaptive Framework: The first integrated application of RTS-inspired $Q$ adaptation combined with innovation-based $R$ tuning for monocular vision-based relative navigation in space.
Real-Time Adaptation: The system adapts to both measurement imperfections (via $R$ ) and dynamic model uncertainties (via $Q$ ) without batch post-processing or prior scheduling.
High-Fidelity Simulation: Implementation of a realistic ENVISAT simulation in Blender with ray-casting visibility checks and synthetic LiDAR generation, surpassing previous MATLAB-based simulations.
LiDAR Bias Estimation: Introduction of a state-augmented bias correction for LiDAR depth, resolving systematic errors between vision and range sensors.

4. Results

The framework was evaluated using extensive Monte Carlo simulations (100 runs) under various conditions:

Accuracy: The dual-adaptive filter ( $MTF + Q_{adaptive}$ $M T F + Q_{a d a pt i v e}$ ) achieved the lowest Root Mean Square Error (RMSE) across all 12 state variables compared to non-adaptive filters, single-adaptation filters, and variational Bayesian methods.
- Example: In eclipse scenarios, the adaptive filter maintained tight error bounds, while the non-adaptive filter showed significant divergence.
Robustness: The system remained stable during measurement outages (eclipses) and at low update rates (0.1 Hz), whereas non-adaptive filters became overconfident (covariance underestimation) and diverged.
Consistency: The Scaled Normalized Estimation Error Squared (SNEES) metric showed that the adaptive filters achieved values close to 1.0 (statistically consistent), whereas non-adaptive filters exhibited overconfidence (SNEES > 1) or excessive conservatism (SNEES < 1).
LiDAR Impact: Adding LiDAR bias estimation further reduced rotational errors and improved filter consistency.

5. Significance

Operational Safety: This work significantly advances the capability for safe proximity operations in ADR missions by ensuring reliable navigation even when visual data is intermittent or noisy.
Autonomy: The elimination of manual tuning and the ability to handle unmodeled dynamics make this framework a strong candidate for onboard, flight-ready deployment.
Future Impact: The methodology paves the way for more complex missions, such as Simultaneous Localization and Mapping (SLAM) for unknown targets and active guidance laws that optimize trajectories based on filter uncertainty.
Next Steps: The authors plan to transition from simulation to Hardware-in-the-Loop (HIL) testing at Iowa State University to validate the framework under physical constraints.

In conclusion, this paper presents a robust, mathematically rigorous, and computationally efficient solution for relative navigation, effectively solving the "tuning problem" in dynamic space environments through dual adaptive noise estimation.