Automation Transformer for Scaled-Down Real-World Control

Abstract

Modern machine learning models may handle particular downstream problems either zero-shot or with limited task-specific datasets to a high degree of performance by transferring information from vast, diversified, task-agnostic datasets. Although this ability has been proven in other disciplines like computer vision, natural language processing, and speech recognition, it has yet to be proven in robotics, where the generalisation abilities of the models are especially crucial due to the challenge of gathering real-world robotic data. We contend that open-ended task-agnostic training paired with high-capacity architectures that can take in all the different robotic data is one of the secrets to the development of such universal robotic models. In this study, we introduce the Robotics Transformer model class, which has potential scalable model features.