In this thesis we would focus on recent developments in diffusion as generative models able to imitate complex data distributions. This can be seen in Stable Diffusion as one of the state-of-the-art Image generators, imitating the originally provided data.
There have already been a few works looking into using similar techniques to imitate robot movements using diffusion models. One of the prominent examples would be Diffusion Policy:
https://diffusion-policy.cs.columbia.edu/You would first review the current state-of-the-art and focus on the theoretical understanding of these models. Next you would look into finding a gap, or an idea to improve these methods, implement these improvements, and apply them so several existing imitation learning environments.
In the end we hope to also create your own imitation learning task on the real robot, record demonstrations for them, and apply your method.
