Neuro-Symbolic Robotics

Abstract

This article introduces and summarizes the emerging field of Neuro-Symbolic Robotics. The advancements in computational power, robust neural structures, and extensive data have positioned Neural Networks as the preferred solution for robotic challenges involving emergent behavior, learning, adaptation, and, more recently, reasoning and communication. Despite these strengths, the deployment of robots in real-world settings demands properties like verifiability, explainability, and interpretability, which Neural Networks lack. Furthermore, neural network-based models experience difficulties with generalization and extrapolation, thus restricting their use. Historically, symbolic systems have been integral to intelligent robotics due to their verifiability, explainability, and scalability, though their manually programmed frameworks fail to manage the complexity and diversity of the robot's continuous and high-dimensional environments effectively. This paper examines various robotic architectures that combine neural networks with symbolic systems in diverse manners to leverage their distinct advantages. We classify these robotic systems into four main categories: intertwined, coupled, non-uniform neuro-robotic systems, and neuro-symbolic translation. We provide an in-depth analysis of the strengths and weaknesses of these systems and outline the future challenges in this domain.