about the location coordination and eval details

[tianbaochou]

We test mimo-embodied with FlagEvalMM frameworks, since mimo-embodied use qwen2.5 as pretrained sources.
It's maybe predict location 2d boxes or points with absolute coords. However, we get 48.98 acc in RoboAfford-Eval (official show 69.81).
Please help us with corrlect prompt or some eval details

[LeiZhouMi]

Thanks for your interest. We use absolute coordinates in our evaluation setup.

Please feel free to check the evaluation script we released in this repo for the exact prompting and evaluation details.