Curriculum Reinforcement Learning for Fuel-Optimal Spacecraft Guidance and Control with Constraints

What is it about?

This work proposes a new reinforcement learning (RL) strategy for training spacecraft guidance and control systems to perform fuel-optimal missions while satisfying mission constraints. The key idea is to use a curriculum learning approach based on the Оµ-constraint method, where constraints (path and terminal) are initially relaxed and then gradually tightened during training. This progressive enforcement allows the RL agent to better explore the solution space before being confined to feasible trajectories, improving both learning stability and final performance. Applied to a simulated pinpoint lunar landing, the method outperforms standard RL techniques by achieving higher constraint satisfaction and fuel efficiency, without relying on precomputed reference trajectories or prior knowledge of the optimal solution.

Featured Image