Training hardware can wait while data is decoded, transformed or copied into the form a model expects. That waiting is part of the experiment’s cost even when it happens outside the optimizer. Robotics makes the problem visible because a training example combines images, actions and state that must stay aligned.
A September 24 engineering account describes LeRobot reading Lance data directly from object storage. In a matched 10,000-step SmolVLA job on eight H100s using DROID, the authors report 1 hour 27 minutes against 2 hours for their standard local-reader configuration, with the same loss. This is a particular system comparison, not a new October 27 robot trial. Read the original source.
Keep sampling equivalent
The important boundary is the data path. A storage integration can change when bytes are fetched and which frames are decoded without changing the intended training examples. A valid comparison needs to preserve the sample sequence and learning setup so that a throughput improvement is not purchased by silently giving the model a different task.
The reported loss agreement is useful for that bounded comparison. It does not establish equal performance on every downstream robotic task, nor does it show that all network environments will behave like the tested setup. Object-store latency, caching and transformations can move the bottleneck elsewhere.
Version the training slice
For practitioners, a reproducible audit starts with a pinned dataset revision and a record of the examples actually consumed. Resume behavior matters too: a job that restarts quickly but repeats or omits data has changed the experiment. The cost of preparing a local copy should remain visible alongside the training run rather than disappearing from the accounting.
The integration also makes curation a question of references and versions. If a team selects a training slice, it needs to retain the rule that produced it and the dataset version to which that rule applied. Otherwise a later run can use a different collection while carrying the same informal experiment name.
Test another workload
The short-term opportunity is to test whether a direct storage reader removes a bottleneck in an existing pipeline. The broader possibility is less duplicated data preparation across experiments. That benefit remains conditional on equivalent sampling, reliable access and an economical balance between computation and data transfer.
The next useful evidence is an independently repeatable run on a different workload with cold-start, restart and downstream checks included. The current result identifies a concrete infrastructure improvement to investigate while keeping the robot’s learned capability a separate measurement.
Source demonstration
The following official-source presentation illustrates the workflow or supplier example discussed above. It is not an independent acceptance test.
References
Questions, news tips or corrections? Contact [email protected].


