Towards Self-reflecting Machines: Two-Minds in One Robot

Abstract

We introduce a technique that allows a robot to increase its resiliency and learning skills by exploiting a process akin to self-reflection. A robot contains two controllers: A pure reactive innate controller, and a reflective controller that can observe, model and control the innate controller. The reflective controller adapts the innate controller without access to the innate controller’s internal state or architecture; Instead, it models it and then synthesizes filters that exploit its existing capabilities for new situations. In this paper we explore a number of scenarios where the innate controller is a recurrent neural network. We demonstrate significant adaptation ability with relatively few physical trials.