Learning new tasks with imitation learning often requires hundreds of demos. Check out our #NeurIPS paper in which we learn new tasks from few demos by inverting them into the latent space of a generative model pre-trained on a set of base tasks.
avivne.github.io/ftl-igm/