Improved Deep Deterministic Policy Gradient for Dynamic Obstacle Avoidance of Mobile Robot
针对传统深度确定性策略梯度算法只能感知固定数量动态障碍物的问题，提出一种结合长短时记忆网络编码器的改进算法，使移动机器人在随机工作场景中实现动态避障，仿真和实验验证了其有效性。
When a mobile robot is required to perform tasks in the unknown and complex environment, it is critical to have the ability of dynamic obstacle avoidance. However, conventional deep deterministic policy gradient (DDPG) for collision-free navigation can only perceive a fixed number of dynamic obstacles, and thus it cannot adapt to the stochastic working scenario. To overcome the limitation, an improved DDPG algorithm is proposed in this study. It is an exploration to implement the DDPG with long short-term memory (LSTM) network-based encoder to achieve dynamic obstacle avoidance for the mobile robot in the stochastic working scenario, which can encode the variable number of obstacles into a fixed-length representation. Specifically, to facilitate the LSTM network-based encoder, one safe processing rule is designed to guarantee the entire information of the observable obstacles can be represented completely. The LSTM network-based encoder takes the latest environment information of observable obstacles by employing the safe processing rule and generates the fixed length state vector. In addition, continuous state space for mobile robot and obstacles, as well as reward function and action space are designed. Both simulations and experiments are conducted and the results verify that the improved DDPG algorithm can achieve collision-free trajectory with multiple dynamic obstacles well. It helps to reduce the path distance and motion time effectively.