youliangtan

/

gr00t-n1.5-libero-spatial-posttrain

like 0

Hi, thank you for releasing the GR00T-N1.5 checkpoints and evaluation code！

We tried to evaluate the Libero Spatial setting using the released checkpoint
youliangtan/gr00t-n1.5-libero-spatial-posttrain, following the instructions in the README.

Specifically, we:

• started the inference service with the provided checkpoint, and
• ran the Libero evaluation script under examples/Libero/eval.

However, in our setup, the evaluation consistently results in 0% success rate across tasks.

Interestingly, when we evaluate our own reproduced checkpoints using the same evaluation pipeline and environment, the results look normal and reasonable. This makes us wonder whether there might be:

• a mismatch between the released checkpoint and the current evaluation configuration,
• some additional setup details or version constraints that we might be missing,
• or a potential issue with the released checkpoint itself.

Could you please let us know:

1. whether this checkpoint has been verified with the current Libero evaluation script, and
2. if there are any specific configuration details (e.g., embodiment-tag, denoising steps, environment versions) that are required for correct evaluation?

Any clarification would be greatly appreciated.
Thanks again for the great work!