Any plans for benchmarking FastSAC and building a standalone version?

[vassil-atn]

Hi,

Thanks for the great work and open-sourcing the repository! It seems that fastSAC works significantly better than PPO for the Humanoid tasks. I'd be curious to see how well it compares on the IsaacLab suite of tasks, is this something planned (like for FastTD3)?

I've also been trying to re-implement it in a standalone way alla rsl_rl_lib so that it can be better integrated with existing IsaacLab environments, but I am getting significantly worse performance on ANYmal locomotion tasks, possibly due to some implementation differences that I'm trying to track down.

Finally, in your opinion, how crucial is the C51 critic to the approach? I've been testing with a HL-Gauss style classifier critic, which was shown to outperform C51 in a recent Deepmind paper, but I do wonder if that might be causing the bad performance.

[SakuBorder]

Hi,

Thanks for the great work and open-sourcing the repository! It seems that fastSAC works significantly better than PPO for the Humanoid tasks. I'd be curious to see how well it compares on the IsaacLab suite of tasks, is this something planned (like for FastTD3)?

I've also been trying to re-implement it in a standalone way alla rsl_rl_lib so that it can be better integrated with existing IsaacLab environments, but I am getting significantly worse performance on ANYmal locomotion tasks, possibly due to some implementation differences that I'm trying to track down.

Finally, in your opinion, how crucial is the C51 critic to the approach? I've been testing with a HL-Gauss style classifier critic, which was shown to outperform C51 in a recent Deepmind paper, but I do wonder if that might be causing the bad performance.

Hello, I attempted to build my implementation based on rsl_rl as well, but the training fails to converge. In addition, the episode length is significantly shorter than that achieved with PPO. Could you please let me know whether your implementation converges successfully?