Abstract:
Modular robot and multi-robot systems share a concept in common: composition,  i.e. the study of how parts can be combined so they can be used to achieve certain objectives. Our vision is to enable robotic systems to configure and reconfigure themselves during field deployment, either autonomously or with the help of users, to adapt to emerging tasks and conditions. This goal requires us to generate compositions in real-time, while maintaining the ability to handle emergent constraints and conflicting objectives. To address these challenges, we present evolution-guided generative adversarial networks (EG-GAN) that learns to map task to compositions. Our method trains a generative model to map a task to a distribution of compositions, with training signals guided by the output of evolutionary algorithm operations. Once trained, the EG-GAN can be used to produce compositions in a near real-time fashion. We demonstrate the effectiveness of our algorithm on two distinct composition problems: 1. designing modular robots and 2. forming teams for multi-robot systems.