Spaces:
Running
Query regarding the performance contribution of Relative vs. RABC in the second fine-tuning stage.
Dear authors,
Thank you for your excellent work. We are currently preparing to adapt this approach to our own robotics tasks.
In reviewing your results, I noticed that the second round of fine-tuning utilizes a combination of both Relative and RABC methods. While the performance gains (from 2.2 to 2.5) are significant, it remains difficult to decouple the specific contribution of each module.
To help us prioritize our optimization efforts for our new task, I was wondering if you might have any additional ablation studies or finer-grained data regarding this second stage? Specifically, have you performed experiments isolating the performance of the model when fine-tuned with only Relative or only RABC?
Any insights into which module proved more critical for the observed performance boost would be invaluable for our own experimental design.