小天体多节点着陆器强化学习任务规划方法

To solve the strong coupling of node motions in the landing process of small celestial bodies which are proposed in China，the complexity of the planning process and the high real-time requirements，a multi-node lander reinforcement learning task planning method is proposed. Firstly，in this method，the state space planning logic model is described as a matrix，and the task planning is modeled as a Markov process by using the model matrix. Then the randomly generated states are filtered through the breadth-first search algorithm based on the state hash to generate the effective multi-node lander state space initial and terminal states. After that，a virtual state space is constructed to train the reinforcement learning agent by randomly switching the initial and terminal states，which improves the planning and adaptation ability of the agent with multiple constraints in task planning. Adaptation ability of the agent are analyzed for the task planning under various conditions. The simulation for landing site detection and analysis task in the landing process is performed，and the experiment shows that the agent can successfully complete all the tests after training，while the planning speed is faster compared with the POPF3 planner. The planning speed advantages in the short-sequence task when there is frequent adjustment，which can be better applied to the task planning of multi-node lander.