docs: HF model card — best-of-both front-matter derived from manifest

README.md

@@ -1,12 +1,22 @@
 ---
-tags:
-  - OpenRAL
-  - rskill
-  - ros2
-  - moveit
-license: apache-2.0
 language:
-  - en
+- en
+license: apache-2.0
+pipeline_tag: robotics
+tags:
+- OpenRAL
+- rskill
+- ros2
+- moveit
+- franka_panda
+- ur5e
+- ur10e
+- so100_follower
+- openarm
+- rizon4
+- sawyer
+- widowx
+inference: false
 ---
 
 # rskill-moveit-eef-pose
@@ -16,7 +26,7 @@ language:
 > end-effector pose goal (position + orientation) through the same
 > `ExecuteRskill` path used by VLA skills. No model weights — the manifest
 > is the entire artefact. New under
-> [ADR-0052](../../docs/adr/0052-moveit-goal-builder-library.md)'s
+> [ADR-0054](../../docs/adr/0054-moveit-goal-builder-library.md)'s
 > `rskill-moveit-*` family.
 
 This package uses `kind: ros_action` (see
@@ -25,7 +35,7 @@ on `RSkillManifest.kind` that selects the
 [`ROSActionRskill`](../../python/rskill/src/openral_rskill/ros_action_rskill.py)
 engine at resolve time, with `ros_integration.goal_builder: pose` selecting
 the [`PoseGoalRskill`](../../python/rskill/src/openral_rskill/pose_goal_rskill.py)
-goal-lowering adapter (ADR-0052). It is the generic Cartesian sibling of
+goal-lowering adapter (ADR-0054). It is the generic Cartesian sibling of
 [`rskill-moveit-look-at`](../rskill-moveit-look-at/) (gaze is a computed-pose
 specialisation) and of [`rskill-moveit-joints`](../rskill-moveit-joints/)
 (joint-space). The engine constructs an `rclpy.action.ActionClient` on the
@@ -111,6 +121,18 @@ of every `ActionChunk` today
 Packing the full trajectory as one chunk with `horizon=N` would let
 waypoints 1..N actuate unchecked.
 
+### GPU-accelerated planning (cuMotion, ADR-0065)
+
+On a host that clears the cuMotion GPU floor (`RobotCapabilities.supports_cumotion()`
+— Ampere+, CUDA ≥ 13, ~8 GB VRAM), the runner sets
+`MotionPlanRequest.pipeline_id = "isaac_ros_cumotion"` so MoveIt plans with
+NVIDIA's CUDA-accelerated cuMotion pipeline; otherwise it falls back to OMPL.
+Transparent — same skill, no manifest change — and it never bypasses the safety
+kernel: the planned trajectory still replays through `/openral/candidate_action`
+and is validated waypoint-by-waypoint. Install: see
+[`docs/contributing/toolchain.md`](../../docs/contributing/toolchain.md) →
+"GPU motion planning — cuMotion".
+
 ## How it was trained / Upstream provenance
 
 Nothing is trained — this rSkill wraps the upstream MoveIt motion planner and
@@ -216,7 +238,7 @@ commercial-use-permissive.
 
 ## See also
 
-- [ADR-0052 — MoveIt goal-builder library + rskill-moveit-* rename](../../docs/adr/0052-moveit-goal-builder-library.md)
+- [ADR-0054 — MoveIt goal-builder library + rskill-moveit-* rename](../../docs/adr/0054-moveit-goal-builder-library.md)
 - [ADR-0024 — ROS-wrapped rSkills](../../docs/adr/0024-ros-wrapped-rskills.md)
 - [`openral_rskill.ros_action_rskill`](../../python/rskill/src/openral_rskill/ros_action_rskill.py) — engine source
 - [`openral_rskill.pose_goal_rskill`](../../python/rskill/src/openral_rskill/pose_goal_rskill.py) — goal-lowering adapter