Skip to content

URML (open robot intent language): intent as a bound on a Brax-trained policy, request for comment #673

Description

@idoco2003

Hi Brax maintainers,

URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: English sentence -> typed primitive -> static validation against a capability manifest and a safety envelope -> dispatch. Brax is interesting to URML from a different angle than a classic simulator: a policy trained in Brax becomes a controller that URML dispatches to, and URML's capability manifest plus envelope statically bound what that learned controller is allowed to attempt.

I will be honest that "learned controller as a URML substrate" is newer ground for the project, so this is genuinely a request for comment.

Two real questions. First, what is the natural boundary between a URML primitive and a Brax environment's action and observation spaces? Second, given Brax's MJX lineage, how do you see the relationship to MuJoCo (we have a separate MuJoCo thread), and does a thin intent/spec layer above a trained policy resonate or seem out of place?

Full write-up, with the manifest mapping table: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0329-brax-outreach.md

One small thing: the GitHub API did not surface an SPDX license id at our verification time. Is Brax Apache-2.0?

Thanks for Brax and for making large-scale differentiable physics approachable.

Ido Yahalomi (URML, greenvh@gmail.com)

AI-assisted prose, maintainer-reviewed before posting (see VIBE.md). Human-only correspondence available on request.

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