Multimodal Robot Systems and Learning

Abstract

In this work we broadly explore the engineering design and system analysis of a multimodal, robotic environment. We first give background on why this type of system is unique, describing the different approach we take to sensing, dynamics, and control. We then delve into the robot itself, and review our development of a python control library enabling a high-level abstraction of low cost hardware. Next we explain the multimodal sensing and physical environment we created for the robot, including some of the initial challenges that forced critical design decisions. Following that, we explain different methods for multimodal representation learning that we tried, and reveal the difficulties we discovered in this task. Finally, we explore some critical takeaways and advocate for a specific path of future work.