As an efficient path planning algorithm, MPPI has received significant attention in the field of path tracking and planning for mobile robots. However, traditional MPPI tends to fall into local minima under extremely clutter environments and thus generates infeasible control sequences. To address this problem, an MPPI control algorithm based on encoder-decoder LSTM is proposed to handle the optimal motion planning problem in clutter environments.
First, the encoder-decoder LSTM network is designed to give MPPI a reference point at each time step. Then, according to the current position and the reference point, the PD controller is used to get the control input at that moment based on the traditional LSTM network framework, which is taken as the sample mean. To improve the real-time performance and computational efficiency of the algorithm, the MPPI sampling part is executed in parallel on the GPU platform, while other control and processing parts are calculated on the CPU.
This paper presents three sets of comparative experiments, including both simulations and physical tests, to evaluate the effectiveness of the proposed algorithm. In challenging environments, the algorithm achieved a hundred percent success rate, successfully guiding the transport robot to move items from the starting point to the target destination while avoiding obstacles along the way.
The algorithm proposed in this paper offers a reference method for robot navigation in complex environments. It addresses the issue of traditional MPPI algorithms being prone to local minima, thereby generating more efficient control sequences.
