ScaRLS: Scalabe Robot Learning System for Robot Manipulation

Abstract

In the field of robotic manipulation, achieving both high generalization capability and high reliability (i.e., high task success rates) remains a central challenge. A critical bottleneck constraining current system performance is the insufficiency of data space coverage—specifically, the robot’s exploration and comprehension of the state space (comprising both observations and actions) are severely limited. To overcome this bottleneck, this paper proposes ScaRLS (Scalable Robotic Learning System), a scalable learning framework designed for robotic manipulation tasks. By leveraging a distributed edge computing architecture, ScaRLS facilitates the acquisition of broader data distributions and enables efficient learning from such data. ScaRLS establishes a complete system closed-loop featuring asynchronous data collection (incorporating both autonomous model inference and human-in-the-loop intervention), asyn- chronous training, and asynchronous parameter distribution. This architecture not only accelerates the workflow of individual computing nodes but also enables large-scale data acquisition and high-efficiency learning by scaling up the number of computing nodes. Furthermore, the system demonstrates robust versatility, offering adaptability to diverse algorithmic models and robotic embodiments.