LeRobot Datasets: October 2025 Update
Note: This article is intended for hobbyists and people familiar with LeRobot — the open-source robotics framework by Hugging Face.
In my previous article I created a baseline view of the LeRobot ecosystem: how many datasets exist, which robots dominate, and what a typical recording setup looks like.
This time, I want to zoom in on October 2025 to get a clearer sense of how quickly things are moving.
How Many Datasets Were Created?
October set a new record with 2,989 new datasets - a 33% increase from September and the highest month ever. That's approximately 96 datasets per day. The community's momentum continues to accelerate.
LeRobot 0.4.0 was released this month together with dataset format v3.0. The adoption was rapid and 1,633 datasets (~54.6%) were created under this new version.
The two biggest improvements in this release for me:
- Native support for NVIDIA Isaac GR00T N1.5 and Pi0.5.
- Ability to stream datasets directly from Hugging Face.
How Large Were LeRobot Datasets in October?
The distribution remains consistent with previous trends - most datasets contain fewer than 30 episodes.
Key observations:
- The majority of datasets have 1-5 episodes, suggesting quick testing and iteration cycles.
- The 50-episode standard: A notable cluster appears at exactly 50 episodes, likely reflecting a community best practice.
- While rare, some datasets exceed 100,000 episodes.
Which Robot Types Were Most Popular in October?
To keep things simple, I highlighted only the top groups and combined the rest into “Other” and even with this, the overall trend is easy to see.
Key observations:
- SO-101 continues to dominate the dataset count, maintaining its position as the most popular robot type.
- BI SO-101 (bimanual version) has appeared in the rankings, showing growing interest in dual-arm setups.
- Interestingly, there's a significant number of datasets with
Nonein therobot_typefield, suggesting some contributors aren't properly setting this metadata.
Here’s a complete list of all robot_type values that appeared in October:
What Was the Typical Camera Setup in October?
The camera setup preferences remain unchanged. Most robots continue using 1-3 cameras, with the 2-camera setup being the most popular.
What Camera Resolutions Were Most Common in October?
Key observations:
- 640×480@30fps remains the gold standard, accounting for 66.5% of all cameras.
- Nearly 20% of cameras use non-standard configurations, showing the community is still exploring optimal setups.
Who Were the Most Active Contributors in October?
Let's take a look at the people driving the LeRobot ecosystem forward this month. These are the top contributors ranked by the number of datasets they created in October 2025:
| Author | Datasets Created | Total Episodes Contributed |
|---|---|---|
| jisoosong | 182 | 952 |
| brucekimrok | 136 | 509 |
| cijerezg | 99 | 1,570 |
| yinxinyuchen | 62 | 620 |
| vlasil | 50 | 1,153 |
| rli14 | 46 | 3,393 |
| 1g0rrr | 46 | 586 |
| lyl472324464 | 45 | 3,584 |
| jskim | 45 | 797 |
| mickael-chen | 42 | 923 |
After clicking through some of the datasets, I noticed that many appear to be copies or variations of the same base dataset, with people experimenting with different settings and configurations. This explains why some contributors have such high dataset counts. They're essentially iterating on the same videos rather than creating entirely new recordings.
Conclusion
October 2025 shows the LeRobot community continues to grow steadily, with familiar patterns holding strong:
- SO-101 remains the most popular robot type, maintaining its position as the go-to choice for hobbyists.
- Camera setups stay consistent, most people continue using 640×480 at 30fps configuration.
- The community stays experimental, small datasets dominate, suggesting people are still exploring and testing rather than committing to large-scale data collection.
If you'd like to see specific metrics tracked more regularly, let me know which ones interest you most!