Skip to content

Return substeps for reward accumulation? #296

@aaprasad

Description

@aaprasad

Currently mjx_env.step(state, action, n_substeps) only returns the final state of substepping. However, this makes it so that rewards can only be obtained for that state instead of being accumulated/averaged over the substeps as is normally done when frame skipping. I think this is a pretty simple feature in theory - just have an arg called return_substeps=False and return all the states over the substeps. I'm just wondering if theres a more memory efficient way to do so?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions