Robust adaptive guidance for autonomous asteroid landing via search-based meta-reinforcement learning

Future asteroid missions require safe landings despite limited prior knowledge and significant uncertainties, posing a critical challenge to current autonomous guidance strategies. This paper introduces a novel robust adaptive guidance framework that integrates meta-reinforcement learning with Monte Carlo Tree Search (MCTS) to enable both rapid learning and efficient adaptation to diverse asteroids. The framework leverages a recurrent network within its meta-reinforcement learning architecture to perceive and respond to dynamic system parameters, ensuring adaptability across varied mission scenarios. The network is trained via an MCTSbased optimization algorithm, where the tree search enhances policy exploration and effectively handles the high-latency rewards of the landing task. Moreover, we introduce an enhanced MCTS by incorporating double progressive widening modifications to refine the deployed action policies. Numerical simulations demonstrate the proposed framework's superior performance and robustness in achieving reliable landing guidance across a wide range of environmental uncertainties.