Question about the RECAP baseline

[xiaojxkevin]

Dear authors,
Thanks for releasing this interesting work.

I have a question about the comparison between RISE and RECAP. The appendix states that the RECAP baseline “follows the recipe of the policy warm-up stage.” My understanding is that this baseline is trained on the fixed offline dataset, while RISE receives additional imagined interactions during self-improvement.

However, the original RECAP algorithm is also iterative: after updating the policy, it collects new real-robot rollouts with the latest policy and retrains on the accumulated data. Therefore, a full RECAP comparison would seemingly also contain new online actions and states, analogous to the final row of Table III, except that the transitions come from the real environment rather than the world model.

Therefore, it would be great if you could clarify:

1. Did the RECAP baseline include any iterative real-robot data collection after policy warm-up?
2. If not, should the reported comparison be interpreted as static RECAP-style warm-up vs. warm-up plus imagined self-improvement, rather than RISE vs. full iterated RECAP?
3. Have you considered comparing RISE against RECAP with a matched number of newly collected real state-action pairs, or alternatively reporting performance under matched real-robot interaction budgets?

This clarification would help me a lot for distinguishing the benefit of world-model imagination from the benefit of simply receiving additional interaction data. And thanks in advance.

BR,
Jinxi Xiao