Proposal: BerkeleyHumanoidLiteAdapter for URML's substrate-neutral robot-intent language (sim via Isaac Lab + real hardware via published Python interface)

[idoco2003]

Hi Hybrid Robotics Lab,

Proposing a BerkeleyHumanoidLiteAdapter targeting both the sim path (Isaac Lab tasks) and the real-hardware path (the published Python control interface) for the Berkeley Humanoid Lite platform. URML Layer-2 primitives (move_to, measure, wait_for, report) map onto Isaac Lab task vectors and whole-body controller setpoints respectively.

URML is an Apache 2.0 specification for substrate-neutral robot intent at urml.dev. Berkeley Humanoid Lite is the most distinctive open-hardware humanoid target available right now: genuinely open (MIT code + CC-BY-SA 4.0 assets), affordable (sub-$5k BOM), research-fresh (2025 release, arXiv 2504.17249), and sim-friendly (Isaac Lab tasks ship in-repo).

The cross-link worth flagging: URML's existing RFC-0050 outreach to NVIDIA Isaac Lab + Isaac-GR00T composes with this proposal directly. A URML-aware Berkeley Humanoid Lite branch where Isaac Lab tasks consume URML primitive sequences closes the loop: train in URML-aware sim, deploy on real hardware via the same adapter family.

This is proposal-only, posted as part of URML's Move #4 outreach. No adapter code in this PR. The sim-vs-real adapter split, the policy-trained-motion primitive mapping (a move_to-to-gait/posture rule similar to URML's existing Petoi RFC-0062), and the authoritative manifest values (DOF / mass / height) are observable choices worth your input before shipping.

Full RFC: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0069-berkeley-humanoid-lite-outreach.md

Feedback we'd value

1. Adapter home. URML repo (reference/humanoid-runtime/src/humanoid_runtime/berkeley_humanoid_lite/), HybridRobotics contributed example, both?
2. Sim-vs-real adapter split. Two adapters (one per substrate) or one adapter with a mode: parameter?
3. Authoritative manifest values. Could you confirm the DOF, mass, height, and control-loop frequency for the production design so URML's manifest reflects ground truth?
4. Policy-trained-motion primitive mapping. Is the move_to → whole-body-setpoint mapping the right shape, or would HybridRobotics recommend a more explicit posture() / gait() Layer-3 vocabulary?
5. Isaac Lab cross-coordination. Interest in coordinating with URML's open RFC-0050 outreach to the NVIDIA Isaac team, given your platform's Isaac Lab task ecosystem?
6. Hardware-in-the-loop conformance path. What is the documented path for a URML conformance run on real hardware?
7. Conformance lane on the README. Open to a URML conformance line on the Berkeley Humanoid Lite README?

Thanks for Berkeley Humanoid Lite and for the genuinely open-hardware posture. The sub-$5k BOM plus Isaac Lab task ecosystem is exactly the substrate-fungible target URML's outreach landscape was missing.

Ido Yahalomi (URML maintainer, urml.dev)