Vision-Based Hazard Detection and Real-Time Safe Trajectory Optimization Using Successive Convexification for Lunar Landing
Author(s)
Brown, Abinay Joel
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
With the renewed interest in returning to the Moon, achieving autonomous and safe lunar descent and landing is paramount for the success of future missions. Safe and autonomous descent also reduces the overhead logistics of pre-planning, surveying, and mapping lunar landing sites. During the descent phase of a lunar mission, the lander must steer away from rough terrain hazards and descend safely to the surface while constrained by fuel, maneuverability, and localization uncertainty. This thesis presents a framework for real-time state estimation, hazard detection from descent imagery, safety-constrained trajectory optimization, and tracking controller for guidance. The proposed framework navigates using dead reckoning and Kalman filter-based sensor fusion to track the descent path using accelerometer and altimetry data while leveraging U-NET image segmentation and contour detection to identify lunar hazards such as craters from cameras that are then scaled and translated for hazard avoidance optimization. A full-horizon, fixed final time, minimum control trajectory optimization problem with control and dynamic constraints is solved using successive convexification (SCvx) optimization scheme. Where the hazard regions identified from descent imagery are represented as ellipse-avoidance chance constraints enforced at the terminal state, ensuring the lander avoids hazards with high confidence despite positional uncertainty. The use of SCvx enables the decomposition of the nonlinear thrust vectoring dynamics and constraints into a series of convex subproblems, significantly enhancing computational efficiency and achieving real-time feasibility. This optimization is performed thrice during descent, allowing the lander to continuously reoptimize and adapt its descent trajectory in response to updated hazard maps generated from the imagery. A Lyapunov-based tracking controller translates the commands from the optimization into thruster and gimbal inputs to track the descent to the safe landing site at a faster update rate. The U-NET segmentation model is trained on synthetic lunar images, the Lyapunov control law is derived to ensure global asymptotic tracking, and the framework is simulated using matplotlib to generate descent imagery as synthetic images.
Sponsor
Date
2025-08-06
Extent
Resource Type
Text
Resource Subtype
Thesis