Hi @lizaijing @MrOK24 , thanks for releasing the work.
I noticed that the reproduced pi0.5 results on CALVIN ABC→D differ between this (BehaviorVLA) and OptimusVLA (https://github.com/iLearn-Lab/CVPR26-OptimusVLA):
BehaviorVLA: 94.8 / 88.9 / 84.1 / 79.7 / 72.1, Avg. Len 4.21
OptimusVLA: 94.4 / 88.4 / 85.3 / 80.1 / 76.1, Avg. Len 4.26
Both are marked as reproduced results, but the CALVIN baseline code and configurations do not appear to be released.
Could you provide detailed CALVIN training settings for pi0.5 baseline? (such as dataset frequency, action horizon, eval replan steps, num training steps)
Thanks!
Hi @lizaijing @MrOK24 , thanks for releasing the work.
I noticed that the reproduced pi0.5 results on CALVIN ABC→D differ between this (BehaviorVLA) and OptimusVLA (https://github.com/iLearn-Lab/CVPR26-OptimusVLA):
BehaviorVLA: 94.8 / 88.9 / 84.1 / 79.7 / 72.1, Avg. Len 4.21
OptimusVLA: 94.4 / 88.4 / 85.3 / 80.1 / 76.1, Avg. Len 4.26
Both are marked as reproduced results, but the CALVIN baseline code and configurations do not appear to be released.
Could you provide detailed CALVIN training settings for pi0.5 baseline? (such as dataset frequency, action horizon, eval replan steps, num training steps)
Thanks!