Dynamic Path Planning for Unmanned Ground Vehicles Based on LSTM and Distributed PPO in Off-Road Environments

Path planning for unmanned ground vehicles (UGVs) in off-road environments faces challenges such as inadequate trafficability assessment, local optima issues, and limited dynamic obstacle avoidance. This paper presents a dynamic path planning algorithm called LSTM-DPPO, which combines long short-term memory (LSTM) and distributed proximal policy optimization (DPPO). Firstly, a UGV trafficability map was created that incorporates three environmental factors, which provides reliable environmental representation for UGV. Secondly, a comprehensive reward function was developed, which combines dynamic obstacle avoidance with trafficability assessment. This function facilitates the collaborative optimization of both path safety and overall trafficability. Additionally, to address the issues of ineffective dynamic obstacle avoidance and the challenges of model convergence commonly found in reinforcement learning algorithms, we introduced an LSTM network to build a dynamic obstacle prediction module. We also employed a distributed PPO architecture to enhance the training speed of the model. Finally, a comparative simulation experiment was conducted between the LSTM-DPPO algorithm and six path planning algorithms, and the generalization ability of the LSTM-DPPO algorithm was verified in three new scenarios. The results indicate that the average inference time of the complete LSTM-DPPO algorithm is 1.74ms, which meets the requirements of realtime path planning. Although its planning time is slightly longer than DPPO algorithm and CV-DPPO algorithm (Constant Velocity Model, CV), it is much faster than the other four algorithms. The LSTM-DPPO algorithm excels in obstacle avoidance, has the highest success rate, and demonstrates good generalization in new scenarios.