Investigation of Stochastic Deep Learning Motion Planning Methods for Autonomous Robots
Effective path planning is an essential part of motion planning for autonomous mobile robots and vehicles. Classical sampling-based path planning methods, such as RRT* are not time-efficient since they must first create an occupancy map before generating a path. Recently, neural network-based planners have been developed to solve this problem, which can quickly generate paths on unseen maps without an occupancy map required. However, these planners perform poorly on unseen maps. To increase success on unseen maps, stochastic elements are included in two new neural network-based planners called Noise, Displacement, Map - GAN (NDM-GAN) and Stochastic-LSTM (S-LSTM). NDM-GAN performs a series of convolutions on a combination of random noise, the start and goal points, and the map, while S-LSTM uses an encoded map and a tensor holding the current and goal points to make a path. Experiments with show that these new planners are successful between 68.58%
- 93.40% of the time. Also, on the unseen maps, NDM-GAN and S-LSTM can generate a path up to 44.7897x and 379.3125x faster than RRT*, respectively. It is also shown that paths generated by NDM-GAN and S-LSTM often possess promising characteristics, such as being shorter than the RRT*-generated paths, and having a larger clearance from obstacles. Since NDM-GAN and S-LSTM are not guaranteed to find a path, if planning reliability is most important, then a classical method like RRT* is preferable.