Strands Robots streams training data from buckets without download

The post walks through a four-stage data loop: record a demonstration into a bucket, store it with byte-level deduplication, train by streaming from the Hub, and deploy the policy back to hardware. Each stage shares a single backend—the Robot() factory that records datasets, syncs them into buckets, streams them back for training, and deploys to physical hardware with a one-keyword-argument change.

In Step 1, the agent records a LeRobotDataset through a DatasetRecorder in the same format LeRobot writes on actual hardware. An example prompt creates a world with the SO-100 robot, adds a red cube and front camera, starts recording with repo_id='local/cube_pick' and fps=30, runs the mock policy for 60 steps, then stops. The sync_dataset_to_bucket() call then moves the finished on-disk dataset into a Storage Bucket at hf://buckets/{bucket}/{run_id}. For hardware recording on an SO-101, LeRobot's lerobot-record CLI handles leader-follower bring-up and writes to disk in the same format, so the same sync call moves it to a bucket: sync_dataset_to_bucket("./recordings", "my-org/robot-fave", run_id="run-021").

Step 2 explains the deduplication savings. Fixed cameras on an arm clearing a table for eight hours record mostly repeated pixels—the same lighting, chassis, background. On versioned repositories, changing one frame in a multi-gigabyte video shard re-uploads the whole file. Buckets are backed by Xet, which uses content-defined chunking: chunk boundaries follow content, so inserting bytes changes only the chunk they land in. Hugging Face's measurements show roughly 4× reduction in bytes transferred per upload across the Hub. A concrete benchmark: starting from 500 MB, changing 1% of bytes moved only 5.5 MB on re-upload; changing 5% moved 27.5 MB; changing 10% moved 55 MB. Without chunk-level deduplication, overwriting an object requires sending all its bytes regardless of what changed.

Step 3 streams the dataset back for training. The same Robot() reads it from the bucket frame by frame, decoding camera video on the fly with no local copy. The whole loop runs in a handful of lines: sim.stream_dataset("my-org/robot-fave/cube_pick", repo_type="bucket").dataloader(batch_size=64) yields batches without downloading the full dataset first. Step 4 deploys the trained checkpoint to the same Robot with one keyword argument change (mode="real" for hardware). Demonstrations recorded on hardware return to the same bucket, closing the loop.

The post notes that LeRobot's format is already used by over 90,000 datasets and models on the Hub from more than 8,000 publishers (LeRobot Project Pulse), so Strands Robots recordings integrate without conversion. Buckets were announced in March 2026 and live in the same hf:// namespace as dataset repositories, using the same hf CLI and permissions users already have. For running on hardware, an SO-101 follower-leader pair or any LeRobot-supported robot works with calibration files under ~/.cache/huggingface/lerobot/calibration/. The minimal path to run the loop on a laptop requires Python 3.12+ on Linux or macOS, a Strands-compatible model provider (Amazon Bedrock, Anthropic API, OpenAI, or local Ollama), and strands-robots[sim-mujoco,lerobot]>=0.5.1; the lerobot extra pulls LeRobot (>=0.6.1), datasets, av, and torchcodec so recording and video decoding work without further setup.