Hi, impressive project!
I have a question regarding the evaluation of Being-H0.7. Since the reported results are achieved with large-scale pretraining on egocentric human videos and robot demonstrations, it is difficult to disentangle how much of the gain comes from the proposed latent world-action architecture itself versus the benefit of large-scale pretraining data.
Would it be possible to share results for Being-H0.7 without large-scale pretraining (for example, training only on benchmark-specific data or using the same training setup as baselines)? I think such an ablation would provide stronger evidence that the performance improvements are due to the model design itself, rather than primarily the scale of pretraining.
This would be very valuable for researchers who want to better understand the architectural contribution of Being-H0.7.
Thanks again for your great work.
Hi, impressive project!
I have a question regarding the evaluation of Being-H0.7. Since the reported results are achieved with large-scale pretraining on egocentric human videos and robot demonstrations, it is difficult to disentangle how much of the gain comes from the proposed latent world-action architecture itself versus the benefit of large-scale pretraining data.
Would it be possible to share results for Being-H0.7 without large-scale pretraining (for example, training only on benchmark-specific data or using the same training setup as baselines)? I think such an ablation would provide stronger evidence that the performance improvements are due to the model design itself, rather than primarily the scale of pretraining.
This would be very valuable for researchers who want to better understand the architectural contribution of Being-H0.7.
Thanks again for your great work.