Hi GELLO community,
URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: a person expresses an intent, URML turns it into a typed primitive, validates it against the robot's declared capabilities and a safety envelope, then dispatches. GELLO is a default way to teleoperate and collect demonstrations, and URML is interesting to a teleop / data-collection rig in two ways that don't compete with manual control.
Nothing here asks the project to adopt, host, or maintain anything. This is a request for comment.
Two seams: (1) shared-autonomy handoff -- the operator stays in control by default, and when they issue a high-level command ("pick that up", "move to the bin") URML validates it against the robot's capabilities + envelope and dispatches, with teleop as the correction path; URML adds a capability/envelope gate a raw teleop stream doesn't have. (2) typed-intent annotation -- each demonstration segment is labelled with the URML primitive it represents (grasp($obj), move_to(bin)), so a recorded demo carries a typed, validatable intent next to the trajectory: structured supervision a downstream policy can learn from, and a record the validator can check against the manifest.
Two real questions: (1) Is a validated shared-autonomy handoff interesting on a GELLO rig? (2) Is labelling demonstration segments with a typed URML primitive useful for the data you collect -- and which is the cleaner first seam?
Full write-up: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0479-gello-outreach.md
Thanks for GELLO; a low-cost, widely-used teleop rig is exactly where this kind of validated-handoff / typed-demo question should be asked.
Ido Yahalomi (URML, greenvh@gmail.com)
AI-assisted prose, maintainer-reviewed before posting (see VIBE.md). Human-only correspondence available on request.
Hi GELLO community,
URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: a person expresses an intent, URML turns it into a typed primitive, validates it against the robot's declared capabilities and a safety envelope, then dispatches. GELLO is a default way to teleoperate and collect demonstrations, and URML is interesting to a teleop / data-collection rig in two ways that don't compete with manual control.
Nothing here asks the project to adopt, host, or maintain anything. This is a request for comment.
Two seams: (1) shared-autonomy handoff -- the operator stays in control by default, and when they issue a high-level command ("pick that up", "move to the bin") URML validates it against the robot's capabilities + envelope and dispatches, with teleop as the correction path; URML adds a capability/envelope gate a raw teleop stream doesn't have. (2) typed-intent annotation -- each demonstration segment is labelled with the URML primitive it represents (grasp($obj), move_to(bin)), so a recorded demo carries a typed, validatable intent next to the trajectory: structured supervision a downstream policy can learn from, and a record the validator can check against the manifest.
Two real questions: (1) Is a validated shared-autonomy handoff interesting on a GELLO rig? (2) Is labelling demonstration segments with a typed URML primitive useful for the data you collect -- and which is the cleaner first seam?
Full write-up: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0479-gello-outreach.md
Thanks for GELLO; a low-cost, widely-used teleop rig is exactly where this kind of validated-handoff / typed-demo question should be asked.
Ido Yahalomi (URML, greenvh@gmail.com)
AI-assisted prose, maintainer-reviewed before posting (see VIBE.md). Human-only correspondence available on request.