Hands-On One-shot Learning with Python

MAML application – domain-adaptive meta-learning

When it comes to imitation learning, robots need to receive proper data consisting of information about kinesthetic changes (awareness about their body parts movements), teleoperation (control), and other kinds of input. On the other hand, a human brain can learn simply by watching some videos. DAML attempted to solve the problem of imitation learning by using meta-learning (MAML). It proposed a system for learning robotic manipulation skills from a single video of human, by just leveraging strong priors (for example, information on kinesthetic learning) extracted through data from different tasks, as shown in the following diagram:

As robots can't be trained using imitation-learning loss ...