Dispatch
AWS and Hugging Face close the robot data loop: record from a natural-language prompt, stream training back at ~1,100 MB/s
Published August 13, 2026, this joint Amazon/Hugging Face writeup wires Strands Agents, LeRobot, and HF Storage Buckets into a single record-store-train-deploy loop with no local dataset copy — camera streams as 200 MB MP4 shards, telemetry as 100 MB Parquet shards, all in unified LeRobotDataset format. Concrete numbers: Xet content-defined chunking cuts repeated uploads ~4x, a single NVIDIA L4 ran 500 ACT policy training steps in ~133 seconds on a 120-frame episode, and warm CDN reads sustained ~1,086 MB/s at 10 GB and ~1,124 MB/s at 100 GB. It targets SO-100/SO-101 arms today via a registry covering arms, humanoids, mobile bases, and hands.
↳ Follow the thread