【摘要】 多机器人系统在救援任务中具有重要意义，本研究以动态灭火任务为应用背景，基于多智能体深度强化学习算法，提出一种能够适应动态未知环境的多机器人协同路径规划算法LA-MASAC。首先，以动态扩散火源为任务模型，引入机器人的动力学方程，在满足障碍物约束与灭火能力约束的条件下，将该问题建模为部分可观察马尔可夫决策过程（POMDP）。使用人工势场法（APF）设计多智能体软演员-评论家（MASAC）算法的奖励函数，并将长短时记忆网络（LSTM）集成到MASAC算法的Actor-Critic结构之中，加快算法的收敛速度，解决了稀疏奖励的问题，提供一种近似最优策略。通过仿真实验，本研究算法有效且具有一定的优越性。更多还原

【Abstract】 Multi-robot systems are of great significance in rescue tasks. Taking dynamic fire-fighting tasks as the application background, a multi-robot collaborative path planning algorithm LA-MASAC based on multi-agent deep reinforcement learning algorithms, which adapts to dynamic unknown environments was proposed in this study. Firstly, with a dynamically spreading fire source as the task model, the robot’s dynamic equations were introduced. Under the constraints of obstacles and fire-extinguishing capabilities, the problem was modeled as a Partially Observable Markov Decision Process(POMDP). The Artificial Potential Field(APF) method was used to design the reward function of the MASAC algorithm, and the Long Short-Term Memory(LSTM) network was integrated into the Actor-Critic structure of MASAC to accelerate the algorithm’s convergence speed, solve the problem of sparse rewards, and provided an approximate optimal strategy. Simulation experiments showed that compared with the MADDPG algorithm, the algorithm in this study has certain effectiveness and superiority.更多还原