Record, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
huggingface ·
This article outlines a streamlined workflow for continuously improving robotic policies by integrating AWS Strands Robots, the LeRobot framework, and Hugging Face Storage Buckets. Rather than repeatedly uploading and downloading large datasets, the system uses Xet-backed storage buckets as a mutable working layer that automatically deduplicates data at the byte level. During operation, an agent records robot demonstrations and syncs them to these buckets, avoiding the storage bloat associated with traditional version-controlled repositories. When it is time to train, the platform streams the dataset directly from the Hub frame by frame, eliminating the need for full local downloads and significantly reducing transfer times. Once a policy is trained, it can be deployed back to physical or simulated hardware with a simple configuration change, allowing newly collected data to feed back into the same bucket. The setup requires minimal dependencies and supports both local simulation and real-world robot deployments like the SO-101. By unifying recording, storage, training, and deployment into a single agent-driven loop, developers can efficiently iterate on robotic models while minimizing infrastructure costs and operational complexity.
huggingface