Probabilistic Path Planning using Obstacle Trajectory Prediction

In this paper, we describe a novel approach for navigation of a robot in a dynamic environment by exploiting the benefits of time series analysis of Long Short Term Memory (LSTM) architectures on obstacle trajectories. Most path planning algorithms consider the instantaneous position of the obstacles while generating the path, resulting in frequent re-planning and extended traversal time. However, moving obstacles more often than not tend to follow certain motion patterns in real life scenarios such as the motion of people on roads, building lobbies, shops, etc. In our algorithm, LSTM-based Dynamic Rapidly-exploring Random Trees star (LD-RRT*), obstacle trajectories are treated as sequences where their future position estimates are incorporated in the state validity checker of planning algorithms to generate an optimal path. Taking the future obstacle motion into account reduces the total traversal time and the frequency of re-planning. We propose the use of an LSTM model for future trajectory prediction and an extension of Rapidly-exploring Random Trees for path planning. Comparisons with other state of the art planners show the effectiveness of our algorithm.