Reproduce “Distribution-Agnostic Robust Trajectory Optimization via Chance-Constrained Reinforcement Learning” on OpenResearch

About this paper

Yashdeep Chaudhary, Roberto Armellin, Harry Holt, Marco Sagliano

A framework named Distribution-Agnostic Robust Trajectory Optimization via Chance-Constrained Reinforcement Learning (RCCRL) was developed to robustify spacecraft trajectories against various uncertainties by combining deterministic optimal control with reinforcement learning. The system demonstrated a 5.50% reduction in upper-tail fuel cost and significantly tighter terminal dispersion in an Earth-Mars transfer, while successfully adapting to different uncertainty distributions and maintaining sub-meter accuracy in a pinpoint rocket landing scenario.

Key insights

• Decoupling nominal trajectory design from robust augmentation allows leveraging established optimal control methods while using RL specifically for handling uncertainties in a structured manner.
• A structured affine closed-loop control policy provides both interpretability for aerospace applications and the ability to systematically adjust the trajectory and contract uncertainty spreads.
• Empirical, sample-based evaluation of probabilistic constraints enables the framework to handle distribution-agnostic uncertainties and highly nonlinear dynamics without restrictive analytical approximations.

Summary by alphaXiv.