Data-Driven and Dynamically Feasible Trajectory Generation for Real-Time Powered Descent Guidance and Robotic Exploration
Name
briden-jbriden-phd-aeroastro-2024-thesis.pdf
Description
Thesis PDF
Size
7.41 MB
Format
Adobe PDF
Checksum (MD5)
82c765eae9cdb2e148a074cce4363955
Author(s)
Briden, Julia
Advisor(s)
Linares, Richard
Marzouk, Youssef
Cauligi, Abhishek
Date Issued
September 2024
Publisher
Massachusetts Institute of Technology
Abstract
Increasingly complex and high-mass planetary missions require autonomous long-horizon trajectory generation to achieve dynamically feasible powered descent guidance. While analytical and indirect methods are computationally efficient, significant simplifications of the dynamics and constraints are required for both problem formulations. Numerical optimization algorithms enable minimum-energy trajectory generation subject to system dynamics and safety constraints but currently remain computationally infeasible on flight-grade processors, taking seconds to minutes to compute a single trajectory. The objective of this dissertation is to develop new algorithms to advance the state of the art in trajectory optimization and planning for autonomous systems. Due to the limited computational abilities of radiation-hardened processors and an increased need for spacecraft and robotic autonomy, specialized algorithms capable of running in realtime constitute enabling technologies for space exploration. Three major contributions are developed in this dissertation. First, a transformer neural network-based algorithm is created to predict the tight constraints that recover the solution and parameter sets for constrained optimization problems. By training on prior runs of the numerical optimization solver, the learned mapping can construct a reduced problem formulation that recovers the optimal solution while reducing runtime by up to an order of magnitude. Second, a method to embed problem-specific information into the neural network training process was developed. By embedding the Lagrangian and Lagrangian gradient merit functions into the training process, neural network-generated control policies are biased toward constraint satisfaction. Third, an autonomous hybrid targeting and guidance algorithm was designed to utilize probabilistic risk maps and numerical optimization to select and navigate to minimum-risk landing sites. Applications in planetary powered descent and landing, as well as rover path planning, are used to benchmark algorithm performance.
MIT Department
Massachusetts Institute of Technology. Department of Aeronautics and Astronautics
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
