Proposal: urml-openpi-bridge with a URML Outputs class following openpi's Inputs/Outputs extension pattern

[idoco2003]

Proposing a urml-openpi-bridge package that hooks into openpi's documented Inputs / Outputs extension pattern. A custom URMLOutputs class (mirroring the LiberoOutputs shape in src/openpi/policies/libero_policy.py) translates the policy's policy.infer(example)["actions"] output into URML primitive calls before any motion reaches the substrate adapter. The bridge keeps the raw action tensor on the return dict, so existing openpi evaluation harnesses see the policy as normal. URML emission rides alongside.

URML is an Apache 2.0 specification for substrate-neutral robot intent. Its Layer-2 primitive vocabulary lets π policies retarget across ROS 2 / PX4 / Isaac / MuJoCo / AUTOSAR Adaptive / OPC UA Robotics by switching URML's substrate adapter, without retraining.

This is proposal-only. No bridge code yet, since the Outputs-class-vs-wrapper question and the upstream-vs-standalone packaging question are worth your input first.

Full RFC with proposed package layout, primitive-to-action mapping, drawbacks, and alternatives:

https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0045-physical-intelligence-openpi.md

Feedback we'd value

1. Extension shape. Is a custom URMLOutputs (mirroring LiberoOutputs) the right hook, or would Pi prefer a wrapper around policy_config.create_trained_policy(...) that intercepts infer(...)?
2. Package home. Upstream into Physical-Intelligence/openpi as a reference URMLOutputs, or a standalone third-party PyPI package (urml-openpi-bridge)?
3. Action-chunk semantics. The proposed move_to collapsing rule assumes contiguous joint-target or end-effector-pose tokens. Safe across π₀, π₀-FAST, π₀.₅?
4. Remote inference. URMLOutputs on the websocket server side or the robot client side?
5. Dataset annotation. openpi consumes LeRobot v3; is a sidecar annotation (matching the LeRobot RFC) the right shape?

Thanks for openpi and for the open-research posture. The Inputs / Outputs pattern made this proposal a lot easier to ground.

Ido Yahalomi (URML maintainer, urml.dev)