Hi ExecuTorch community,
URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: an intent becomes a typed primitive, validated against the robot's declared capabilities and a safety envelope, then dispatched. ExecuTorch runs PyTorch models on-device, including the edge compute a robot carries to run a learned policy, and URML is interesting one layer above the inference runtime.
Nothing here asks the project to adopt, host, or maintain anything. This is a request for comment.
The layering: a policy deployed to a robot's edge compute runs under ExecuTorch; the action it produces is then checked by URML against the robot's declared capabilities + the active safety envelope before dispatch. ExecuTorch is the runtime that computes the action; URML is the typed gate that decides whether to dispatch it. The model ExecuTorch runs has the obs/action spaces and training domain a URML "LearnedPolicy" declaration records, so the on-device policy carries the bounds the gate enforces.
Two real questions: (1) is "ExecuTorch runs the policy on-device, URML validates the action before dispatch" a sensible description of the layering for a robot? (2) Is a LearnedPolicy envelope traveling with an on-device model useful for the robotics-deployment case -- and is an inference runtime the right altitude to engage?
Full write-up: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0516-executorch-outreach.md
Thanks for ExecuTorch; on-device policy inference is exactly the do layer a validated-action gate wants to sit above.
Ido Yahalomi (URML, greenvh@gmail.com)
AI-assisted prose, maintainer-reviewed before posting (see https://github.com/URML-MARS/URML/blob/main/VIBE.md). Human-only correspondence available on request.
Hi ExecuTorch community,
URML (urml.dev) is a small, Apache-2.0 language for describing robot intent: an intent becomes a typed primitive, validated against the robot's declared capabilities and a safety envelope, then dispatched. ExecuTorch runs PyTorch models on-device, including the edge compute a robot carries to run a learned policy, and URML is interesting one layer above the inference runtime.
Nothing here asks the project to adopt, host, or maintain anything. This is a request for comment.
The layering: a policy deployed to a robot's edge compute runs under ExecuTorch; the action it produces is then checked by URML against the robot's declared capabilities + the active safety envelope before dispatch. ExecuTorch is the runtime that computes the action; URML is the typed gate that decides whether to dispatch it. The model ExecuTorch runs has the obs/action spaces and training domain a URML "LearnedPolicy" declaration records, so the on-device policy carries the bounds the gate enforces.
Two real questions: (1) is "ExecuTorch runs the policy on-device, URML validates the action before dispatch" a sensible description of the layering for a robot? (2) Is a LearnedPolicy envelope traveling with an on-device model useful for the robotics-deployment case -- and is an inference runtime the right altitude to engage?
Full write-up: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0516-executorch-outreach.md
Thanks for ExecuTorch; on-device policy inference is exactly the do layer a validated-action gate wants to sit above.
Ido Yahalomi (URML, greenvh@gmail.com)
AI-assisted prose, maintainer-reviewed before posting (see https://github.com/URML-MARS/URML/blob/main/VIBE.md). Human-only correspondence available on request.