LeRobot dataset catalog
LeRobot datasets: browse the LeRobot-format catalog
LeRobot datasets are robot-learning datasets published in the LeRobotDataset format on the Hugging Face Hub: Parquet tables for states and actions, MP4 for camera frames, and JSON metadata for episode boundaries. Hundreds are public — BridgeData, DROID, RT-1/fractal, ALOHA sim, PushT — but licenses and embodiments vary, so match each set to your task before you train.
Comparison
| Dataset | What it is | Embodiment | License | Best for |
|---|---|---|---|---|
| bridge_orig_lerobot | LeRobot port of BridgeData V2 (60,096 trajectories, 13 skills, 24 environments) | WidowX arm | Apache-2.0 | Language-conditioned manipulation pretraining |
| droid_lerobot | LeRobot port of DROID (76k demos, 564 scenes, 86 tasks) | Franka Panda | Apache-2.0 | Visual diversity, in-the-wild manipulation |
| fractal20220817_data_lerobot | LeRobot port of the RT-1 “fractal” demonstrations | Everyday Robots mobile manipulator | Apache-2.0 | Large-scale multi-task pretraining |
| aloha_sim_transfer_cube_human | ALOHA simulation, human-teleop cube transfer | Bimanual ALOHA (sim) | MIT | ACT / bimanual policy prototyping |
| pusht | 2D T-block pushing benchmark | Planar pusher (sim) | MIT | Diffusion Policy baselines, quick sanity checks |
| L2D | yaak-ai Learning-to-Drive multimodal driving set | Passenger vehicle | Apache-2.0 | End-to-end / VLA-for-driving |
What “LeRobot datasets” actually means
The phrase collapses two different things. One is the LeRobotDataset format — the on-disk layout a dataset uses. The other is the Hugging Face Hub ecosystem of datasets published in that format. When someone says "I trained on LeRobot datasets," they mean data in the format, pulled from the Hub, not one canonical corpus.
The format is worth understanding before you browse, because it decides how a set loads. LeRobotDataset v3.0 splits into three pillars: low-dimensional, high-frequency signals (states, actions, timestamps) in Apache Parquet; camera frames concatenated into MP4 and sharded per camera; and JSON/Parquet metadata (`meta/info.json` for schema, `meta/stats.json` for normalization, `meta/tasks.jsonl` for language, `meta/episodes/` for offsets) that reconstructs episode boundaries from within shared files rather than from filenames [1]. That decoupling — few large files on disk, intuitive episode-level access in the API — is what lets the format scale to millions of episodes.
[2]"LeRobot aims to provide models, datasets, and tools for real-world robotics in PyTorch."
The practical consequence for a browse hub: two datasets can both be "LeRobot datasets" and still differ on rights, robot, task quality, and schema completeness. Treat the format as a delivery contract, not a quality guarantee. The definitional companion — what LeRobot datasets are — covers the format ecosystem in depth; this page is the browse layer on top of it.
The LeRobot datasets worth knowing
Most of the high-signal public LeRobot sets fall into two groups. The first is Open X-Embodiment ports — community re-hosts of the datasets that already anchor VLA pretraining, converted to LeRobotDataset so they load without a bespoke pipeline. `bridge_orig_lerobot` is the LeRobot port of BridgeData V2's 60,096 language-labeled WidowX trajectories across 13 skills and 24 environments [3]; `droid_lerobot` is DROID's 76k Franka demonstrations across 564 scenes and 86 tasks [4]; `fractal20220817_data_lerobot` is the RT-1 “fractal” corpus from Google's Everyday Robots fleet. All three trace back to Open X-Embodiment's 1M-trajectory, 22-embodiment pool [5].
The second group is LeRobot's own reference datasets — small, clean sets built to exercise specific policies. `aloha_sim_transfer_cube_human` gives you bimanual ALOHA teleoperation in simulation for ACT-style training; `pusht` is the 2D T-block pushing benchmark that Diffusion Policy popularized, ideal for a fast sanity check before you commit compute. Outside manipulation, `L2D` from yaak-ai extends the format to autonomous driving. The table below is the by-task, by-embodiment view; the cluster links further down go straight to each dataset's profile.
How to pick a LeRobot dataset for your task
Format compatibility is the easy part — every set here loads through `LeRobotDataset`. The hard part is fit, and four checks catch most mismatches before they cost you a training run.
- 01
License first, not last
Apache-2.0 and MIT sets (bridge, droid, fractal, ALOHA sim, PushT) are safe for commercial model training. CC BY-NC sets are research-only — a noncommercial license blocks a deployed product regardless of how good the data is. Read the repo card, not the format.
- 02
Match the embodiment
A WidowX policy will not transfer clean to a Franka without action-space re-normalization. Pick a set whose robot matches yours, or budget for the cross-embodiment gap. Sim sets (ALOHA sim, PushT) are for algorithm work, not deployment.
- 03
Check language coverage
VLA training needs instruction-labeled episodes. BridgeData is densely language-labeled; many raw teleoperation sets are not. If your model is instruction-conditioned, filter for language coverage before counting trajectories.
- 04
Verify schema completeness
Confirm the observation keys, action space, and FPS you need are actually present in `meta/info.json`, and that normalization stats exist in `meta/stats.json`. A set can be in-format and still be missing the wrist camera or force channel your policy expects.
Visualizing, streaming, and converting LeRobot datasets
Every LeRobot dataset on the Hub gets a dataset viewer — you can inspect episodes, replay video, and read the schema in the browser before downloading a byte, which is the fastest way to sanity-check task quality. For training at scale, `StreamingLeRobotDataset` consumes a set directly from the Hub without a full local download, so you can start iterating on a multi-hundred-gigabyte corpus in minutes [1].
If you hold your own data, the migration path is short. A converter aggregates the per-episode Parquet and MP4 files that v2.1 produced into the larger shards v3.0 expects and writes the episode offsets into `meta/episodes/` — one command, `convert_dataset_v21_to_v30 --repo-id=<HF_USER/DATASET_ID>` [1]. One gotcha worth internalizing: when you record or build a v3 dataset you must call `dataset.finalize()` before `push_to_hub()`, or the Parquet writers never close their footers and the files load corrupt [1]. For the format-to-format mechanics — RLDS to LeRobot and back — the RLDS-and-LeRobot formats guide walks the conversion end to end [6].
When the public catalog runs out: custom LeRobot-format capture
Public LeRobot datasets cover the tasks the research community happened to record. The moment your spec diverges — a specific gripper, a factory environment nobody has filmed, a task distribution the public sets under-sample, or commercial rights a CC BY-NC set can't grant — the catalog stops helping. That gap is what custom capture fills.
Truelabel is a physical-AI data marketplace: you post a spec, vetted capture partners return sample packets, and accepted batches ship in the format your pipeline already speaks. Teleoperation traces, egocentric video, and manipulation demonstrations come from around 10,000 consented collectors across 100 countries, delivered in RLDS, LeRobot, or MCAP to S3, GCS, or Azure, with per-trajectory provenance and consent artifacts attached. Because delivery is LeRobot-native, the data loads through the same `LeRobotDataset` path as everything above — no conversion project between you and training. Start with an eval-sized request to validate quality, then scale the spec that works.
Related pages
Use these to move from category-level context into specific task, dataset, format, and comparison detail.
External references and source context
- LeRobot dataset documentation
LeRobotDataset v3.0 stores tabular signals (states, actions, timestamps) in Apache Parquet, camera frames in MP4 sharded per camera, and schema/episode-segmentation metadata in JSON/Parquet; v3 packs many episodes per file where v2 used one file per episode, and StreamingLeRobotDataset consumes datasets directly from the Hub. A converter migrates v2.1 datasets to v3.0.
Hugging Face ↩ - LeRobot GitHub repository
LeRobot provides models, datasets, and tools for real-world robotics in PyTorch, and the Hugging Face Hub is the distribution layer for LeRobot-format datasets.
GitHub ↩ - Project site
BridgeData V2 reports 60,096 trajectories (50,365 teleoperated demonstrations plus 9,731 scripted pick-and-place rollouts) across 13 skills and 24 environments, with natural-language instructions, which is why the LeRobot port is a strong language-conditioned pretraining base.
rail-berkeley.github.io ↩ - Project site
DROID reports 76k demonstrations / 350 hours across 564 scenes and 86 tasks, and reports improved policy performance, robustness, and generalization from real-world manipulation data.
droid-dataset.github.io ↩ - Project site
Open X-Embodiment pools more than 1M real robot trajectories across 22 distinct embodiments, and its component datasets (BridgeData, DROID, RT-1/fractal) are the sets most often re-hosted in LeRobot format.
robotics-transformer-x.github.io ↩ - LeRobot documentation
LeRobot documentation describes the dataset tooling, viewer, and training integrations that make the format a practical distribution layer.
Hugging Face ↩
FAQ
What is the LeRobot dataset format?
LeRobotDataset is a standardized format for robot-learning data on the Hugging Face Hub. In v3.0 it stores states, actions, and timestamps in Apache Parquet, camera frames in MP4 (sharded per camera), and schema plus episode-segmentation metadata in JSON/Parquet. Episode boundaries are resolved through metadata, not filenames, so many episodes pack into each file and the format scales to millions of episodes.
Where do I find LeRobot datasets?
On the Hugging Face Hub, tagged with the LeRobot library. The strongest public sets are Open X-Embodiment ports — bridge_orig_lerobot (BridgeData V2), droid_lerobot (DROID), fractal20220817_data_lerobot (RT-1) — plus LeRobot's own reference sets like aloha_sim_transfer_cube and pusht. Each has a browser-based dataset viewer so you can inspect episodes before downloading.
Can I use LeRobot datasets commercially?
It depends entirely on the per-dataset license, not the format. BridgeData, DROID, and RT-1 ports are Apache-2.0 and the ALOHA-sim and PushT sets are MIT, all of which allow commercial model training. Sets under CC BY-NC are noncommercial and cannot train a deployed product. Always read the repository card before assuming Hub availability equals commercial rights.
What changed between LeRobotDataset v2.1 and v3.0?
v2.1 wrote one Parquet and one MP4 file per episode. v3.0 concatenates many episodes into larger shards and reconstructs episode views from metadata, which cuts file-system pressure and speeds initialization at scale. v3.0 also adds Hub-native streaming via StreamingLeRobotDataset. A converter, convert_dataset_v21_to_v30, migrates existing v2.1 datasets in one command.
How do I visualize a LeRobot dataset?
Every LeRobot dataset on the Hugging Face Hub has a built-in dataset viewer that replays episode video and exposes the schema in the browser, so you can check task quality before downloading. For larger sets, StreamingLeRobotDataset lets you iterate on the data directly from the Hub without a full local download.
How do I convert my own data to LeRobot format?
Record with LeRobot's tooling, or convert an existing dataset. To migrate a v2.1 dataset run convert_dataset_v21_to_v30 with your Hub repo id. When building a v3 dataset from scratch, call dataset.finalize() before push_to_hub() so the Parquet writers close their footers — skip it and the files load corrupt. For RLDS interconversion, see the RLDS-and-LeRobot formats guide.
Can I get custom data delivered in LeRobot format?
Yes. Truelabel delivers teleoperation, egocentric, and manipulation data in LeRobot, RLDS, or MCAP to S3, GCS, or Azure. You post a spec, review a sample packet before scale, and receive rights-cleared batches with per-trajectory provenance and consent artifacts — captured to your embodiment and task rather than reused from a research-licensed public set.
Looking for LeRobot datasets?
Specify modality, task, environment, rights, and delivery format. Truelabel matches you with vetted capture partners and helps scope consent artifacts and commercial licensing requirements before delivery.
Request custom LeRobot-format capture