RoBERT: Low-Cost Bi-Directional Sequence Model for Flexible Robot Behavior Control

Abstract

Requirement of human involvement for data collection or system design has always been a major challenge for building robot control policy. In this paper, we present $\textbf{Ro}$bot-$\textbf{BERT}$ (RoBERT), a method to build general robot control policy for complex behaviors with $\textit{least}$ human effort. Starting from unsupervisedly-collected dataset, RoBERT has no requirements of human labels, high-quality behavior dataset or accurate information of system model, in contrast to most other methods for building general robot agent. RoBERT is further pre-trained via $\textit{Masked Action-Inverse-Inference}$ (MAII), a method inspired by $\textit{Masked Language Modeling}$ (MLM) in BERT-like language models and has potential to enable $\textit{zero-shot}$, $\textit{multi-task}$, $\textit{keyframe-based}$ robot control with little architectural change and user-friendly interface. In our empirical study, RoBERT is successfully applied on various types of robots in simulated environment and could generate stable and flexible behaviors to fulfill complex commands.