Data efficiency in imitation learning with a focus on object manipulation

Imitation is a natural human behaviour that helps us learn new skills. Modelling this behaviour in robots, however, has many challenges. This thesis investigates the challenge of handling the expert demonstrations in an efficient way, so as to minimise the number of demonstrations required for robots to learn. To achieve this, it focuses on demonstration data efficiency at various steps of the imitation process. Specifically, it presents new methodologies that offer ways to acquire, augment and combine demonstrations in order to improve the overall imitation process. Firstly, the thesis explores an inexpensive and non-intrusive way of acquiring dexterous human demonstrations. Human hand actions are quite complex, especially when they involve object manipulation. The proposed framework tackles this by using a camera to capture the hand information and then retargeting it to a dexterous hand model. It does this by combining inverse kinematics with stochastic optimisation. The demonstrations collected with this framework can then be used in the imitation process. Secondly, the thesis presents a novel way to apply data augmentation to demonstrations. The main difficulty of augmenting demonstrations is that their trajectorial nature can make them unsuccessful. Whilst previous works require additional knowledge about the task or demonstrations to achieve this, this method performs augmentation automatically. To do this, it introduces a correction network that corrects the augmentations based on the distribution of the original experts. Lastly, the thesis investigates data efficiency in a multi-task scenario where it additionally proposes a data combination method. Its aim is to automatically divide a set of tasks into sub-behaviours. Contrary to previous works, it does this without any additional knowledge about the tasks. To achieve this, it uses both task-specific and shareable modules. This minimises negative transfer and allows for the method to be applied to various task sets with different commonalities.Open Acces