Where does GELLO teleoperation's lasting value sit as egocentric collection scales?

[jaikr-dev]

Hi, and thanks again for the paper, the dataset, and for being open about the roadmap.

In the launch coverage you lay out a three tier data pyramid:

• teleoperation on the deployment robot,
• GELLO teleoperation for more general data,
• and egocentric wearable collection of everyday human tasks.

I'm trying to understand where the GELLO tier sits over the long run. ABC itself shows broad pretraining helps a lot while action grounded data still matters for the eventual task. Here I have a couple of questions:

1. As you scale egocentric collection, where do you see GELLO's durable value? Is it mainly the action fidelity and real robot grounding that egocentric can't yet match, or are there other reasons it
   stays essential?
2. Do you see the tiers as complementary, egocentric for breadth in pretraining and teleoperation for the action grounded finetuning, or do you expect egocentric to gradually absorb the middle tier?

• Jai