Prediction of human behaviour using imitation learning

Summary

This thesis explores the use of the Info-GAIL algorithm, which is based on the generative adversarial imitation learning framework to model modalities of human behaviour towards performing tasks. The goal of this thesis is to use behaviour models learnt through Info-GAIL to predict the modality of executing a specific “object grasping” task. This is done through learning sub-task policies from unsegmented demonstrations of tasks. Specifically, this thesis uses a dataset with trajectories regarding human behaviour towards grasping objects of different sizes in specific . These are pre-processed to correct imperfections and exploited to extract features of trajectory states that are used during training. Then, the implemented method is tested and evaluated utilizing the extracted features. The thesis concludes with a thorough presentation of results and proposals for further work towards using multi-modal imitation learning to predict human behaviour in executing tasks.