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.
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.