Self-Supervised Deep Visuomotor Learning from Motor Unit Feedback

Despite recent success in a number of domains with deep learning, expensive data collection and the need for large datasets becomes a major drawback for deep learning with real robotic platforms. As a result, many of the successful work in deep learning has been limited to domains where large datasets are readily available or easily collected. To address this issue, we leverage closed-loop controllers to reduce the dimensionality of output neurons and introduce a method of self-supervision derived from a dynamical systems approach that captures the robot’s evaluation of its own interactions with the world. We focus on the accurate prediction of controller dynamics (or motor unit status) given visual observation and demonstrate that a real robot platform, the uBot-6, can quickly acquire visuomotor skills that are general enough to be applied to a set of novel objects.