Open Source Robotics

A curated collection of high-quality open source robotics projects, tools, and software to propel the robotics community forward.

LeRobot can now train straight from LanceDB tables in object storage

  • Source: Hugging Face and LanceDB
  • Author: Open Source Robotics
  • Published: 2026-09-24
  • Category: research

LeRobot's dataset layer now reads LanceDB tables natively, so training can stream from object storage with a full global shuffle instead of downloading first. On all of DROID the remote run finished 33 minutes faster than a local copy, with identical loss. The support is on LeRobot's main branch but not yet in a release.

Correction — 29 October 2026. An earlier version of this story gave DROID as roughly 27.6 million frames. The linked post does not state that figure, so it is removed.

---

LeRobot can now train directly from datasets stored as LanceDB tables, reading only the bytes each batch needs from object storage instead of downloading the dataset first. Engineers from Hugging Face and LanceDB described the integration on 24 September, a month after the support itself merged into LeRobot's main branch on 28 August. Lance is the first alternative storage format LeRobot's dataset layer reads natively, and nothing in recording or training code has to change: the existing lerobot-train command points at a Lance table and picks the right loader itself.

The practical effect is that a dataset too large to copy onto every training machine can stay where it is. Pointed at a Hugging Face Storage Bucket, a SmolVLA training run on two H100s consumed the same 212 samples a second from the bucket as it did from a local NVMe copy, with an identical loss curve. The larger test was all of DROID — 369 GB of video — read from a remote object store with a global shuffle over every frame. Across two otherwise identical 10,000-step runs on eight H100s, the LanceDB run finished in 1 hour 27 minutes against 2 hours for the standard reader on a local copy that first needed a 384 GB download. The standard run spent 37 percent of its time waiting for data; the LanceDB run spent under 2 percent. Final loss matched to four decimal places.

The second half of the argument is about curation. Because the training data is already a LanceDB table, embeddings, derived scores and full-text indexes can live as columns on the same rows the trainer reads, so searching and filtering the data does not mean exporting it into a separate vector database. The authors embedded 610,403 DROID frames with SigLIP2 and used combined visual and SQL queries to mine behaviour that had no label — frames of fabric being moved in different homes whose task text never mentioned it. Episodes can be inspected in Foxglove straight from the table.

One full scan of DROID turned up three findings worth knowing whatever you train with: the idle filter in OpenPI removes 11.66 percent of the dataset and hard-codes the resulting length rather than computing it; the default 80/20 split places every building and every data collector in both training and validation sets, so it leaks; and successful episodes turned out to be jerkier than failed ones, which means the common smoothness filter would throw away good data. On LIBERO, where success can be measured in simulation, deliberately corrupting 30 percent of the demonstrations and then removing them with three table queries recovered a better policy than training on the messy set.

Two things to be clear about. The native reader is on LeRobot's main branch but not yet in a tagged release — the latest, v0.6.1, predates it — so today it means installing from source. And the step that adds a computed column without rewriting the data, which underpins the curation workflow, is described as a LanceDB Enterprise feature; the open-source reader and the Lance format are what LeRobot itself depends on. The authors have published an interactive demo running these queries against the full DROID table.

Why it matters

Robot-learning datasets have outgrown the habit of copying them onto every training machine: DROID alone is several hundred gigabytes of video. Streaming from object storage without losing GPU throughput or shuffle quality removes that copy step, and keeping search, scores and embeddings on the same table the trainer reads turns data curation into a query rather than an export pipeline. The findings from one scan of DROID — a leaking default split, a quality filter pointing the wrong way — show why curation is worth making cheap.

Key takeaways

  • LeRobot now reads datasets stored as LanceDB tables natively; recording and training code are unchanged and lerobot-train picks the loader itself
  • Training can stream from a Hugging Face Storage Bucket or other object store with a global shuffle over every frame and no download step
  • On all of DROID, a remote LanceDB run finished in 1 h 27 m against 2 h for a local-copy run, with identical loss and 1.7% of time spent waiting for data instead of 37.4%
  • Embeddings, scores and indexes can live as columns on the training table, so curation and mining are queries rather than exports
  • One scan of DROID found a leaking default train/validation split and a smoothness filter that removes good data
  • The support is on LeRobot's main branch but not yet in a tagged release, and the column backfill used for curation is a LanceDB Enterprise feature

Open-source angle

The integration is in LeRobot itself under its Apache-2.0 licence, merged into the main branch on 28 August, and the Lance format it reads is open source. The line to watch is between that and LanceDB's commercial offering: adding a computed column to a table without rewriting it, the step the curation workflow is built on, is presented as a LanceDB Enterprise feature. Training from Lance tables and querying them does not depend on it.

Related technologies

  • LeRobot
  • LanceDB
  • Lance
  • SmolVLA
  • DROID
  • Foxglove
  • Hugging Face Storage Buckets

Links

  • Source: https://huggingface.co/blog/CarolinePascal/how-to-train-your-robot-the-lancedb-edition

This website requires JavaScript to function properly. Please enable JavaScript to explore our collection of open source robotics projects.