Local Planning for Unmanned Ground Vehicles using Imitation Learning

Sammanfattning: Mobile robotics is an expanding field worldwide leading to the need for advanced path-planning algorithms that can traverse various environments. Current state-ofthe- art path-planning algorithms used at the Swedish Defence Research Agency, FOI, tend to be inflexible and parameter dependent. The parameters might need to be tuned for each new environment, which is a very labor-intensive process. This thesis investigates the possibility to replace computationally heavy pathplanning algorithms with neural networks using Imitation Learning. Neural networks with and without Long Short-Term Memory (LSTM) layers were trained and evaluated. The network without LSTM failed to capture the temporal dependency of the input data, which lead to poor performance. Using LSTM layers performed close to the imitated algorithm in the training environment and in certain situations, the trained neural network outperformed the algorithm by a big margin. In conclusion, neural networks are, after training, able to replace path-planning algorithms and in certain scenarios, the network outperforms the algorithm. Further work is needed to get a robust local planner with a neural network as a base.