We analyze human involvement in generic imitation learning settings, especially in the data collection phase. We also review methods that aim to reduce such human efforts in the training loop, by either inferring learning signals from poor demonstrations, or extracting latent skills from noisy datasets.