Discrepancy Between Training and Play Performance in Isaac Sim (SKRL PPO)

[RodsCoimbra]

Hi,
I’ve noticed a significant difference between how my agent behaves during training (train.py) and how it performs when playing back the trained policy. During training, the videos show the agents consistently reaching the target. However, when I play back the policy using play.py, the performance is completely different—the agents fail to reach the target at all.

Training:
https://github.com/user-attachments/assets/ffb492e8-2c4a-4a12-afba-2ddd48d0c8d6

Playing:
https://github.com/user-attachments/assets/49d8f4c5-59f7-44bc-8fa8-97da556cd8bb

I’m training with SKRL’s PPO in Isaac Sim 4.5 + Isaac Lab 2.0.
Any ideas on what might be causing it or if it can be a bug?
Thanks!

[RandomOakForest]

Thank you for posting this. I will move this post to our Discussions section for the team to follow up.

[giangdao1402]

Have you figure out the way to solve the problem, i encountered the same one @RodsCoimbra