Hi, thank you for releasing the work.
Have you tested GQS with recent motion-tracking controllers such as SONIC or HoloMotion?
I tried applying GQS to the released SONIC repository using the BoneSeed dataset, evaluating both the 3% and 10% subsets. Visualization confirmed that Stage 1 filters out many visibly problematic motions, but I could not reproduce the reported result that the selected subset outperforms the full dataset.
I also tested GQS with an independent SONIC reimplementation for another humanoid robot, this time using AMASS, and again could not reproduce the improvement.
Have you observed the 3% or 10% result with SONIC, HoloMotion, or other recent trackers? Any details about the selected motion lists, preprocessing, and training configuration would be helpful.
Hi, thank you for releasing the work.
Have you tested GQS with recent motion-tracking controllers such as SONIC or HoloMotion?
I tried applying GQS to the released SONIC repository using the BoneSeed dataset, evaluating both the 3% and 10% subsets. Visualization confirmed that Stage 1 filters out many visibly problematic motions, but I could not reproduce the reported result that the selected subset outperforms the full dataset.
I also tested GQS with an independent SONIC reimplementation for another humanoid robot, this time using AMASS, and again could not reproduce the improvement.
Have you observed the 3% or 10% result with SONIC, HoloMotion, or other recent trackers? Any details about the selected motion lists, preprocessing, and training configuration would be helpful.