Stochastic learning and extremal-field map based autonomous guidance of low-thrust spacecraft

Area of Science:

• Aerospace Engineering
• Control Theory
• Machine Learning

Background:

• Designing autonomous guidance laws for low-thrust spacecraft is complex, especially for time- and fuel-optimal trajectories.
• Existing methods often require significant computational resources or lack adaptability to real-world conditions.

Purpose of the Study:

• To develop an autonomous guidance law for low-thrust spacecraft using Gaussian Process Regression (GPR).
• To address both time-optimal and fuel-optimal trajectory design challenges.
• To create a robust guidance system applicable to interplanetary transfers with potential off-nominal performance.

Main Methods:

• Employed a supervised stochastic learning method, Gaussian Process Regression (GPR).
• Utilized a "perturbed back-propagation" approach to generate optimal control data for training.
• Trained the GPR model on an "extremal bundle" of neighboring optimal trajectories.
• Predicted costate n-tuple or primer vector for time- and fuel-optimal cases, respectively.

Main Results:

• Demonstrated the effectiveness of GPR in designing an autonomous guidance law.
• Successfully applied the methodology to a time-optimal interplanetary transfer (Earth-3671 Dionysus).
• Evaluated the guidance law's performance under various design parameters and off-nominal thruster conditions.
• Confirmed the utility and applicability of the proposed framework.

Conclusions:

• The GPR-based autonomous guidance law is effective for low-thrust spacecraft.
• The "perturbed back-propagation" method provides a viable approach for generating training data.
• The framework shows promise for real-world interplanetary missions, with potential for future enhancements.