Heuristic dynamic path planning algorithm based on SALSTM-DDPG

Abstract

In the field of path planning, there are many excellent traditional algorithms proposed by scholars, but usually these traditional methods based on static environment, these methods lack the ability to deal with dynamic variable environment. A heuristic reinforcement learning dynamic environment path planning algorithm (SALSTM-DDPG) combining self-attention and long short-term memory network(LSTM) is proposed in this paper. Firstly, an automatic encoder is constructed to reduce the feature dimension of the local cost map, which reduces the complexity of the overall model. Then the local cost map and global path after dimensionality reduction are taken as input, the original source of information is guaranteed to the maximum extent possible by this method, at the same time, the global optimal path is used to guide the local path planning. Finally, the deep reinforcement learning algorithm DDPG is used for specific motion control, the Actor part is built using a network based on the SALSTM algorithm. This approach allows previous sequence to be used as a reference when an Actor makes decisions, this enables dynamic obstacles to be predictably avoided by the robot. The feasibility and efficiency of the proposed algorithm are proved by experiments.