URML (open robot intent language): a validated intent layer above ManiSkill policies — request for comment

[idoco2003]

Hi ManiSkill community,

URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: it validates a request against a capability manifest and a safety envelope, then dispatches. ManiSkill is interesting to URML as a high-throughput manipulation benchmark — a clean place to show URML wrapping a learned policy in a validated envelope and to evaluate validated-intent dispatch at scale.

Nothing here asks the project to adopt, host, or maintain anything. This is a request for comment.

A URML intent declares the goal and the envelope; a policy trained in ManiSkill produces the low-level action, and URML validates the request against the declared task capabilities before the policy acts (decide-then-do applied to learning). URML drives a ManiSkill task through its Python interface; URML's optional validation block records the simulation-fidelity context a run was checked in. Validate-before-actuate refuses an out-of-capability request before the simulated robot moves.

Two real questions: (1) Is a validated intent layer + envelope above ManiSkill policies interesting for the manipulation-learning community? (2) What should a URML capability manifest declare to describe a ManiSkill task robot honestly — arm/drive type, reach/DOF, gripper + graspable classes, workspace bounds, observation/action assumptions?

Full write-up: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0449-maniskill-outreach.md

Thanks for ManiSkill; a fast open manipulation benchmark is a great place to think about safety around learned control.

Ido Yahalomi (URML, greenvh@gmail.com)

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

[StoneT2000]

Sorry, I am not really going to respond to a bot. From reading the details here I don't think ManiSkill would benefit from additional hierarchy like this.