Investigation of stochastic deep learning path planning methods for mobile robots

Path planning is essential for mobile robot navigation, especially in complex environments. Traditional methods like RRT* (Rapidly Exploring Random Tree) explore known spaces effectively but lack time efficiency. Recent neural planners generate paths quickly on unseen maps, though often unreliably. This work proposes two stochastic neural planners: Noise, Displacement, Map-Generative Adversarial Network (NDM-GAN) and Stochastic-Long Short-Term Memory (S-LSTM) which integrate structured randomness to enhance generalization. NDM-GAN uses convolutions over random noise, start/goal points, and map data and S-LSTM leverages dropout in latent map-encoded LSTMs. Tested on unseen maps, they achieve up to 93.40% success and generate paths up to 13,793.18% faster than RRT*, with shorter lengths and greater obstacle clearance. Compared to similar planners, they show a 28.3% gain in viable path rates. While not probabilistically complete, these models demonstrate the power of stochasticity in neural planning, offering a strong basis for further work. • Presents two stochastic global path planners for 2D maps using GANs and LSTMs. • Fastest proposed planner is up to 13,793% faster than the classical RRT*. • Stochastic planners generate paths on unseen maps, often shorter than training paths.