One-Shot Imitation from Observing Humans via Domain-Adaptive Meta-Learning

Demonstrations provide a descriptive medium for specifying robotic tasks. Prior work has shown that robots can acquire a range of complex skills through demonstration, such as table tennis (Mülling et al., 2013), lane following (Pomerleau, 1989), pouring water (Pastor et al., 2009), drawer opening (Rana et al., 2017), and multi-stage manipulation tasks (Zhang et al., 2018). However, the most effective methods for robot imitation differ significantly from how humans and animals might imitate behaviors: while robots typically need to receive demonstrations in the form of kinesthetic teaching (Pastor et al., 2011; Akgun et al., 2012) or teleoperation (Calinon et al., 2009; Rahmatizadeh et al., 2017; Zhang et al., 2018), humans and animals can acquire the gist of a behavior simply by watching someone else. In fact, we can adapt to variations in morphology, context, and task details effortlessly, compensating for whatever domain shift may be present and recovering a skill that we can use in new situations (Brass & Heyes, 2005). Additionally, we can do this from a very small number of demonstrations, often only one. How can we endow robots with the same ability to learn behaviors from raw third person observations of human demonstrators?