Clarify 16 supported embodiments vs 7 trained embodiments in released GR00T-H weights

[Hemanth21k]

📚 The doc issue

Hi, I am testing the released GR00T-H weights with Open-H Embodiment datasets and noticed a potential ambiguity between the GitHub documentation, code registry, and Hugging Face model card.

The GitHub README/documentation appears to describe support for 16 Open-H embodiments. However, the Hugging Face model card for nvidia/GR00T-H says:

GR00T-H was trained on 7 different robotic embodiments contained within Open-H, including CMR Versius, dVRK, dVRK-Si, UR5, Rob Surgical Bitrack, Tuodao MA2000, and KUKA.

The released Hugging Face weights also appear to contain mixed metadata:

• embodiment_id.json includes many registered embodiment IDs, including embodiments beyond the 7 listed in the model card.
• processor_config.json and statistics.json do not appear to contain entries for all registered IDs.
• Direct inference fails for at least one code-registered embodiment when its processor/statistics metadata is missing.

For example, TUM_SONATA_FRANKA exists in the codebase and in embodiment_id.json:

• EmbodimentTag.TUM_SONATA_FRANKA
• open_h/embodiments/tum_sonata_franka/tum_sonata_franka_config.py
• embodiment_id.json: "tum_sonata_franka": 7

But the released processor_config.json and statistics.json do not include tum_sonata_franka, so inference fails with:

KeyError: 'tum_sonata_franka'

Suggest a potential alternative/fix

No response