Rearrangement tasks, which involve moving objects from their initial positions to desired target positions, require the coordination of multiple robotic skills. Training a policy to perform such tasks in partially unknown environments is challenging and often requires significant engineering effort. One promising solution is to leverage human demonstrations to learn effective task-execution strategies.

The aim of this thesis is therefore to investigate how recent advances in imitation learning and multi-modal generative models can be used to make efficient use of demonstration data. The learned policies should enable a robot to perform rearrangement tasks and adapt its behavior to new task objectives.