Training Loss Curve

The plot below compares the training loss of:

1. pi05 baseline (red)
2. GroundSG (green)
3. FrameSamp+Modul (blue)

After 80k iterations, the expected training losses are approximately 0.003, 0.0025, and 0.0015 for pi05 baseline, GroundSG, and FrameSamp+Modul, respectively. The loss curves of other MME-VLA variants are similar.