Data diffusion and data format conversion

[FALCONYU]

1. Can datasets collected in Lerobot format be disseminated?
2. Can data formats between different Lerobot versions be converted? I noticed that the data format collected in version 0.2.0 is different from the latest data format.
   Thank you!

[stbzwyd]

同问，现在是V3，openpi都是V2搞的。有没有什么办法。。。

[Littlew69]

Me too! May I know the exact format of episodes_stats.jsonl? I have no idea about it.

[tc-huang]

Hello @FALCONYU,

About data version conversion, this merged PR 1412 and this doc mention the method for conversion from v2.1 to v3.0, and this open PR 2109 is about converting back from v3.0 to v2.1.
I also found this older merged PR 711 which seems to have mentioned the method for conversion from v2.0 to v2.1, but the file used (convert_dataset_v20_to_v21.py) looks like it might have been removed.
I'm not sure if this is what you are looking for. Since I haven't needed to do a conversion myself, I haven't actually tried out the methods in these PRs. Hope this helps!

[FALCONYU]

@tc-huang Awesome！thanks a lot, this solved my second problem.

[FALCONYU]

@stbzwyd @Littlew69 I have tested it, the version conversion problem, this script is available.You can download this script directly without reinstalling Lerobot. When I ran this script, the Lerobot version in my environment was the latest 0.3.4.

https://github.com/bingogome/lerobot/blob/v3_0-to-v2_1/src/lerobot/datasets/v30/convert_dataset_v30_to_v21.py

[tc-huang]

@FALCONYU, could you please clarify what you mean by 'disseminated' in your first question?

[FALCONYU]

@tc-huang I'd like to perform data diffusion on Lerobot-formatted data collected in a real-world environment, similar to NVIDIA's IsaacLab Mimic. This method allows for data diffusion of both video and motion data, allowing for more data to be obtained from a limited dataset. However, IsaacLab Mimic uses the HDF5 format for data diffusion. Therefore, I'd like to find a method to directly perform data diffusion on Lerobot-formatted datasets, as the number of datasets I actually collect is very limited.

[FALCONYU]

@tc-huang我想对在真实环境中收集的 Lerobot 格式数据进行数据扩散，类似于 NVIDIA 的 IsaacLab Mimic。这种方法可以同时对视频和运动数据进行数据扩散，从而可以从有限的数据集中获取更多数据。然而，IsaacLab Mimic 使用 HDF5 格式进行数据扩散。因此，由于我实际收集的数据集数量非常有限，我想找到一种直接对 Lerobot 格式数据集进行数据扩散的方法。

@Littlew69
I can convert an HDF5 dataset to Lerobot V2 format, but I'm not sure how to reverse the process. Furthermore, the diffusion process in IsaacLab Mimic probably requires a simulation scenario, while I'm collecting data in a real-world environment. Therefore, I'm unsure if I can achieve my goal using IsaacLab Mimic through some kind of conversion.

[tc-huang]

@FALCONYU, I'm afraid I don't have the relevant experience to help.

[FALCONYU]

@tc-huang Ok, thanks, I'll keep looking for a solution

[bingogome]

Hey @tc-huang @FALCONYU , I have a conversion from v2.1 to v2.0 here. Works well for me. I wrote both the PRs for (probably) exactly the same use case as you are looking at.

[tc-huang]

@bingogome, thanks for sharing!