Trained checkpoint perform weird while evaluation

[HeinrichvonAnonym]

Hello, developers,
I trained a checkpoint of the task Franka reaching with the official code, and the training converged successfully.
$ python reaching_franka_isaacgym_skrl_train.py

But when I evaluate the checkpoint with this checkpoint (I have copied the trained checkpoint to the relative path and renamed it),
$ python reaching_franka_isaacgym_skrl_eval.py
the performance was like an initial model.

But the the ideal performance is obtained when I instead use the official checkpoint download from https://skrl.readthedocs.io/en/latest/intro/examples.html#real-world-examples.

It seems that the problem is not by the player or the environment, but the training process seems to be successful, I can't understand this result, can anybody help to find where the issue is? thank you!

GPU: RTX 4090
Driver: 570
Cuda: 12.2
python: 3.7