What happened?

In a blog post published on August 13, 2026, Hugging Face explained how AWS's open-source Strands Robots SDK is being combined with the LeRobot data format and Hugging Face Storage Buckets. The post shows how the process—from recording robot demonstrations to training to deploying on hardware—can be run within a single agent loop.

Strands Robots is a library that abstracts robot arms, humanoids, mobile bases, and hands, exposing them to Strands agents as tools (AgentTool) alongside the LeRobot stack. The SDK is distributed under the Apache 2.0 license and builds on the Robot() factory introduced in an earlier blog post. This factory works by calling one of many supported robot types—arm, humanoid, mobile base, hand—by name.

Why does it matter?

LeRobot's dataset format is used across more than 90,000 datasets and models from over 8,000 publishers on the Hugging Face Hub. This scale means a dataset recorded with Strands Robots can be read directly with existing LeRobot tools without any conversion step.

According to the blog post, the real problem arises when this loop needs to run continuously: as episodes recorded throughout the day grow, every training run copies the entire dataset to GPUs, every new checkpoint is pushed out, and the next batch of recordings comes back in. These repeated operations lead to the same bytes being transferred over and over.

Hugging Face Storage Buckets, announced in March 2026, is presented as the answer to this problem. This Xet-powered, unversioned, mutable storage type sits alongside dataset repositories under the hf:// namespace and works with the existing hf command-line interface (CLI). It forms the working layer between the moment of recording and the moment of training.

How does the new system work?

  • Recording: The Robot() object records a LeRobotDataset and syncs the completed dataset to a Storage Bucket; each sync uploads only the changed bytes.
  • Training: The dataset is read frame by frame via streaming from the Hugging Face Hub without being downloaded; camera videos are decoded on the fly during this process.
  • Deployment: The trained checkpoint is sent to physical hardware through the same Robot() object, simply by changing a single keyword argument (mode="real").
  • Format consistency: At every stage of the process, the data remains in the on-disk format written by LeRobot; no conversion step is required.

What's needed for setup?

According to documentation published by Hugging Face, the minimum requirements to run the system include Python 3.12 or higher, Linux or macOS (with MuJoCo support for Apple Silicon), and a Strands-compatible model provider such as Amazon Bedrock, the Anthropic API, OpenAI, or local Ollama. Installing Strands Robots together with its dataset extras (the lerobot extra) automatically brings in the LeRobot, datasets, av, and torchcodec libraries.

For hardware and advanced use, a Hugging Face account with write access and a token, the hf CLI, calibration files along with an SO-101 follower-leader robot pair or another LeRobot-supported robot are required; additionally, an NVIDIA GPU is needed for local vision-language-action (VLA) inference, or a GPU cluster for training at scale.

What's next?

The blog post notes that a runnable example of the loop described is shared in the examples/notebooks/05_streaming_data_loop.ipynb file. In the post's default configuration, a mock policy is used instead of a real one—producing valid but not particularly useful datasets—while the lerobot[training] package must be installed separately for the actual training step.