Cagatay Cali

Collecting robot data is still the slow part. You record episodes, wait on the upload, wait again for the download onto whatever GPU you managed to get, and then find out the scene drifted a third of the way through and half of it is junk. Full post with code examples: huggingface.co/blog/amazon/s… This walkthrough with AWS Strands Agents closes that loop. An agent drives the SO-101 - in sim or on the real arm - and records episodes directly in LeRobot format, then syncs them to a Hugging Face Storage Bucket. Xet dedups every push at the byte level, so re-syncing a dataset where 1% of the content changed moves 5.5MB instead of the full 500MB. Training reads straight out of the bucket with StreamingLeRobotDataset: MP4 shards and Parquet state-action files pulled as the loader needs them. No download step, the GPU starts on step one. 500 ACT steps in 133s on a single L4. Deploying back to the arm is mode="real". Same dataset, same format, from the first recorded episode to the policy running on hardware. The agent is what makes this a loop instead of a pipeline. It decides which episodes are worth keeping, notices when the scene drifts, checks whether there's enough data to train on yet, and picks which checkpoint goes to the robot. Swap the provider name for SmolVLA, Pi0, GR00T or Cosmos.