Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets

Why it matters

This article is a technical walkthrough from AWS and Hugging Face showing how a continuous robot-learning data loop can run inside a single Strands agent. It builds on a prior post introducing Strands Robots, an open source SDK that exposes robot abstractions, simulation, and the LeRobot stack as agent tools. Here, the focus shifts to the data side: recordings are synced into Hugging Face Storage Buckets, a mutable, non-versioned object-storage repository type that sits in the same namespace as dataset repositories. The key efficiency argument is that repeated transfers of the same bytes become costly over a long collection campaign, and the described loop addresses this by uploading only changed bytes per sync and streaming datasets directly from the Hub for training rather than downloading them. The same Robot object both records and reads back data, and a trained checkpoint deploys to hardware with a single keyword argument change. The setup spans simulation and physical robots such as the SO-100 and SO-101, with LeRobot's widely adopted dataset format ensuring compatibility.