Amazon's open-source Strands Robots SDK allows automated agents to manage the entire robotics data loop from a single process, according to technical details published by Hugging Face on August 13, 2026. The Apache 2.0-licensed library connects simulation, telemetry recording, policy training, and physical hardware deployment into one unified system.
The workflow relies on Hugging Face Storage Buckets, a mutable object-storage repository type introduced in March 2026. Storage Buckets sit beside standard dataset repositories in the Hugging Face namespace but accept mutable in-place uploads without creating a new Git commit revision for every individual recording session.
Data deduplication
To prevent repeated network transfer costs during daily collection campaigns, Storage Buckets use byte-level deduplication powered by Xet content-defined chunking. Hugging Face measurements show content-defined chunking reduces data transferred per upload by about four times across the Hub. In benchmark tests on a 500 MB base file, modifying 1 percent of the data required uploading only 5.5 MB, while a 10 percent change moved 55 MB.
Data storage follows LeRobot's default file structures, which are already used across more than 90,000 datasets and models on the Hub from over 8,000 publishers. Sensor telemetry and joint actions write to Parquet shards capped at 100 MB, while camera streams output to MP4 shards capped at 200 MB. Syncing an updated dataset uploads only the newly written or partially filled trailing shards.
Streaming and training benchmark
Model training streams data directly from the Hub rather than copying full datasets to local graphics processing units (GPUs). Parquet files supply joint state vectors while MP4 video shards decode on the fly during training iteration.
In a performance test on an Amazon EC2 g6.4xlarge instance equipped with an NVIDIA L4 GPU, training an Action Chunking Transformer policy with 51.6 million parameters over a 120-frame demonstration completed 500 optimizer steps in 133 seconds. Benchmark testing on an m5dn.24xlarge instance in the us-east-1 region demonstrated warm content delivery network read speeds of 1,086 MB/s on a 10 GB payload and 1,124 MB/s on a 100 GB payload, compared to 780 MB/s for cold reads.
Hardware deployment
Deploying a trained policy to physical arms, such as the SO-101 follower and leader setup, requires changing the mode parameter from simulation to real execution within the same code. Physical demonstrations collected on hardware write back to the local disk in the identical LeRobot format before syncing back to the bucket.
For edge devices running without video decoding libraries, the SDK provides a setting to skip video processing entirely during proprioceptive-only data streaming. Hugging Face confirmed that Storage Buckets currently offer data hosting in United States and European Union regions on enterprise plans, with expansion to Asia-Pacific and Gulf Cooperation Council regions planned.
