source policy training

[HaodongHuang-W]

Your work is exceptionally well done, truly fantastic! I would like to inquire if the source policy can be trained directly using train.py? During the training process, I encountered an issue where the reward function remains at 0, and I noticed that the task registration name in train.py does not match the registration name of the source policy.

[WhoKnowsssss]

Hi there,

Thanks for your kind complements! Unfortunately the source policy training environments are a little bit tricky to integrate. If you'd like to train bipedal walking, you can directly use this repo: https://github.com/HybridRobotics/SymmLoco and use the checkpoint in our repo. The source policy performance will be plausible.

If you'd like to train other skills, please refer to https://github.com/Alescontrela/AMP_for_hardware, though I'd encourage you to try some newer RL frameworks, as our source policies trained with this AMP pipeline was not that great....

Let me know if I could provide further help.

[generatives]

I have been trying to train the source RL policies as well. I am having trouble with the different rsl_rl version that exist in the repo. It seems like they are two different versions, and the envs that are defined in legged_gym are maybe designed for the rsl_rl in the other implementation? It is hard to say.

Are the policies trainable with the repo in it's current state? Do I just need to reformat the configs to match what OnPolicyRunner in legged_gym is expecting?