low success rate when setting goal_cond=True during policy learning

[Yiyao-Ma]

Thank you for this outstanding work.
I have a question regarding the dexgrasp_policy. When I set goal_cond=False using your provided code and data, I can train a model with a high success rate (around 0.8+). However, when I set goal_cond=True, meaning I want the grasp not only to successfully lift the object but also to be similar to the input grasp configuration, the success rate drops significantly (almost to 0). This issue persists even when training on just one object, which is quite strange. Could you advise me on how to address this?