[Question] Performance Difference rsl_rl vs skrl

[j-hae1]

Question

Why do rsl_rl and skrl show a lot difference at their performance? (when training with PPO)

I have changed skrl's loss function (as rsl_rl did, which update scheduler at every mini batch, and compute value function based on maximum difference), but it still shows performance gap between two different libraries.

Is there any specific setting for rsl_rl or skrl?

I just trained at the anymal-C (both manager based and direct rl env ), at the both env the skrl is worse than rsl_rl.

the two upper charts are SKRL
the two lower charts are rsl_rl

the skrl's error has larger value than rsl_rl's error

[RandomOakForest]

Thank you for posting this. I'll move this to our Discussions section for the team to follow up. @Toni-SM for vis.

[Toni-SM]

Hi @j-hae1

Could you please provide or share the changes you did on skrl code and the agent configuration (if you have changed it) for the given tasks (also, please name the task by their full name e.g.: Isaac-Anymal-XXX for a better identification) for investigation?