LeRobotDataset v3.0: 大規模ロボットデータセット対応へ `LeRobotDataset:v3.0`: Bringing large-scale datasets to `lerobot`

Hugging Face has released LeRobotDataset v3.0, a substantial overhaul of the on-disk format used by its open-source robotics library lerobot. The redesign targets a concrete pain point: scaling to datasets with hundreds of thousands of episodes, such as DROID with its 95k+ trajectories, where the previous v2.1 layout simply broke down.

Under v2.1, each episode was stored as its own Parquet file with separate per-camera MP4 videos. That made small datasets easy to inspect but produced file-count explosions that strained both local filesystems and the Hugging Face Hub's Git/LFS backend. Listing, cloning, or even uploading such repositories became painfully slow, and downstream training pipelines paid the cost in metadata overhead.

v3.0 replaces this with a chunked layout. Multiple episodes are concatenated into shared Parquet files, and videos are merged into longer MP4 segments, all grouped into chunk directories with a configurable size limit. A new episodes/ metadata table records, for each episode, which chunk file it lives in, the byte and frame offsets, and timestamp ranges. Random access by episode index thus requires reading only the relevant shard, while sequential streaming becomes far more I/O efficient.

The second headline feature is streaming. The new StreamingLeRobotDataset class lets users iterate over a dataset directly from the Hub without downloading it in full, fetching only the chunks needed for the current batch. This mirrors the streaming mode in the broader Hugging Face datasets library but is tuned for the mixed video-plus-tensor nature of robotics data. For researchers working with multi-terabyte corpora on bandwidth- or disk-constrained machines, it is a meaningful unlock.

The release lands in the middle of an accelerating push toward robot foundation models. Physical Intelligence's Pi0, Google DeepMind's RT-2 and the Open X-Embodiment collaboration, and NVIDIA's GR00T effort all bet that scaling laws extend into manipulation and embodied control, which in turn requires common, web-scale data formats. lerobot has positioned itself as the open-source hub for this ecosystem, and a format that can comfortably hold DROID-class datasets is arguably a prerequisite for that role.

For existing users, Hugging Face provides a conversion script from v2.1 to v3.0, so most migrations should be straightforward. Projects with custom data loaders will need to adapt to the new metadata schema, particularly around episode-to-chunk lookups and timestamp handling. It also seems plausible that future iterations will extend the format toward richer multimodal annotations, language instructions, and perhaps standardized action spaces, though the current release focuses squarely on scalability and streaming rather than semantic enrichment.

Overall, v3.0 is less a flashy feature drop than a piece of infrastructure work that quietly raises the ceiling on what lerobot users can train on. Whether it becomes the de facto interchange format for open robotics data will depend on adoption by dataset publishers, but the technical direction — chunked shards plus first-class streaming — aligns with where large-scale ML data tooling has been converging for years.