Performance Cliffing for G1 Locomotion

[Buzz-Beater]

Hi,

Thanks for the great work. I was trying to reproduce the results from the paper on the G1 locomotion task. However, I observe a significant performance cliff at around 4K steps (figures attached) when I extend the training steps to 5K. Any ideas on why this is happening?

To fully reproduce my issue, I'm training with 4096 environments on a single NVIDIA 3090 GPU using the isaaclab_experiments codebase. All other hyperparameters and package versions match the default description.

Looking forward to thoughts from you guys :-)

[stergiosba]

Hi, did you ever figure this out?