Probabilistic regression for autonomous terrain relative navigation via multi-modal feature learning.

Area of Science:

• Aerospace Engineering
• Computer Science
• Robotics

Background:

• Human spaceflight expansion necessitates advanced autonomous navigation.
• Existing deep learning localization methods for planetary landing have limitations.
• Robustness in diverse scenarios is critical for real-world mission success.

Purpose of the Study:

• To develop and evaluate a novel deep learning framework for spacecraft powered descent navigation.
• To enhance the accuracy and robustness of planetary lander localization.
• To address the need for improved autonomous navigation in complex space missions.

Main Methods:

• A novel formulation for training Convolutional Neural Network (CNN)-based Deep Learning (DL) models in a multi-layer cascading architecture.
• Utilizing classification probabilities from DL models as regression weights for position estimation.
• Leveraging multi-sensor data, including image intensity and depth, for relative spacecraft localization.
• Validation through Monte Carlo analysis across various simulated scenarios.

Main Results:

• Demonstrated efficacy of the proposed DL architecture and state-estimation framework.
• Accurate determination of spacecraft location relative to terrain at specific altitudes.
• Successful navigation performance validated across multiple simulated environments.
• Significant promise for multi-modal feature learning in realistic space missions.

Conclusions:

• The proposed deep learning approach offers a robust solution for autonomous spacecraft landing navigation.
• The multi-layer cascading CNN architecture effectively utilizes multi-sensor data for precise localization.
• This framework shows great potential for enhancing the safety and success of future planetary missions.