AI Pose Estimation in Orbital Satellite Servicing

Orbital servicing spacecraft rely on advanced artificial intelligence for rendezvous pose estimation to autonomously capture and refuel aging satellites in geostationary orbit. By utilizing computer vision and deep learning algorithms, a servicer can determine the exact position and orientation of a target satellite that is tumbling and lacks active communication or navigation aids. This article explores how onboard sensors gather visual data, how neural networks process this information to predict movement, and how robotic systems execute the final capture and refueling maneuvers without human intervention.

The Challenge of Uncooperative Targets

Satellites in geostationary orbit (GEO) that have run out of fuel or suffered system failures often lose their ability to maintain orientation. Without stabilizing thrusters, external forces cause these satellites to tumble unpredictably. These are classified as "uncooperative" targets because they do not have working transponders, telemetry data, or specialized optical reflectors to help a servicing spacecraft dock with them. Furthermore, the immense distance to GEO introduces communication delays that make real-time remote control from Earth unsafe for complex docking maneuvers. Therefore, the servicing spacecraft must rely entirely on onboard autonomy.

Sensor Data Collection

To approach a tumbling satellite, the servicing spacecraft uses a suite of electro-optical sensors, including visible-light cameras, infrared sensors, and LiDAR. These sensors capture continuous real-time data of the target against the harsh lighting conditions of space, where blinding sunlight and pitch-black shadows exist simultaneously. LiDAR provides 3D point clouds to measure distance and shape, while optical and infrared cameras capture structural details regardless of the lighting environment.

AI-Driven Pose Estimation

Once the sensors collect the visual data, it is fed into onboard AI algorithms—typically Convolutional Neural Networks (CNNs)—trained to perform 6-Degrees-of-Freedom (6DoF) pose estimation. Pose estimation is the process of determining the target's three-dimensional position (X, Y, Z coordinates) and its three-dimensional orientation (roll, pitch, and yaw) relative to the servicer.

Because obtaining real images of satellites in space for training AI is difficult, these neural networks are trained on millions of synthetic images generated from 3D models of the target satellite. When the AI views the real satellite, it identifies specific geometric keypoints, such as the corners of solar panels, thruster nozzles, or communication dishes. By matching these detected keypoints to a known 3D wireframe model of the satellite, the AI calculates the target's exact pose in milliseconds.

Motion Prediction and Trajectory Matching

Determining the pose in a single frame is not enough; the servicer must understand how the satellite is tumbling over time. The AI feeds its continuous pose estimations into navigation filters, such as an Extended Kalman Filter. This system analyzes the sequence of poses to calculate the target's angular velocity and axis of rotation.

Once the AI accurately predicts the future motion of the tumbling satellite, the servicer's Guidance, Navigation, and Control (GNC) system takes over. The servicer fires its thrusters to enter a synchronized orbit, matching the exact spin rate and trajectory of the target. To the servicer's cameras, the tumbling satellite now appears stationary, creating a safe window for physical interaction.

Autonomous Grappling and Refueling

With the relative motion neutralized, the AI shifts its focus to fine-guidance tracking of a specific grappling fixture, which is usually the target satellite's marman ring (the structural ring used to attach the satellite to its launch vehicle) or its main liquid apogee engine nozzle.

Guided by real-time pose updates from the AI, a robotic arm extends from the servicer and clamps onto the target. Once a rigid connection is established, the servicer slowly fires its thrusters to halt the tumbling motion, stabilizing both spacecraft. Following stabilization, a secondary robotic tool approaches the target’s fuel valves. It uses localized computer vision to cut through thermal blankets, unscrew the fuel caps, connect a high-pressure hose, and transfer propellant, successfully extending the life of the satellite.