DIFFERENTIABLE ROBOTICS: COMPOSITIONAL DEEP LEARNING WITH DIFFERENTIABLE ALGORITHM NETWORKS

Abstract

Towards human-level robot intelligence a central question is the architecture of robot learning: how is the system represented, and how is it trained?

This thesis introduces the Differentiable Algorithm Network (DAN), a compositional architecture for designing robot learning systems. The DAN is composed of neural network modules, each encoding an algorithm and associated models; and it is trained end-to-end from data. The key idea is to make model-based algorithms differentiable and encode them in a generalized neural network, thus combining the model-based modular system design with data-driven end-to-end learning. Algorithms act as structural assumptions to moderate data requirements; end-to-end learning allows modules to adapt and compensate for imperfections.

We introduce DANs for a range of domains: particle filter networks for visual localization; differentiable SLAM networks for visual SLAM; QMDP networks for partially observable planning. We also combine DAN modules for visual navigation in simulation and with a real-world quadruped robot.