robotiq_description: command the gripper opening in metres

[ebarnett3]

The shipped 2F descriptions model the driven joint as the revolute knuckle, so gripper_cmd goals and /joint_states are in radians against gripper_closed_position. That is the wrong parameterization for a parallel gripper: what callers want is the jaw opening, and every consumer that needs it converts radians back to millimetres on its own.

Parallel grippers are normally modelled as a slider, with the linkage as passive joints for visuals. Franka's hand is the reference shape — finger_joint1 is type="prismatic" in metres, finger_joint2 is a <passive_joint> in the SRDF, and its group states are metres. This description is revolute because it was modelled on the visible knuckle rather than on the useful quantity.

Why it matters

• The MCP server carries its own radians<->mm map (mcp/gripper_mcp/units.py), whose docstring already calls itself a stopgap. It cannot even own the calibration: robot_description.py reads the command joint and gripper_closed_position out of the driver's URDF at runtime so the two agree on "closed".
• The tsf_demo consumers work in mm end to end — grip-width thresholds, opening as a model feature for object recognition — and bypass the ROS driver entirely, opening the serial port with pyRobotiqGripper to call position_mm().
• That makes three definitions of one conversion: the driver's, the MCP's, and pyRobotiqGripper's.

Plan

1. Read side, additive, no URDF change. Publish the opening in metres on the GripperStatus message from Publish object_status (gOBJ) and motor_current so clients can read them #74. This is what the MCP actually needs and it needs no description work at all; it can land first and on its own.
2. Command side. Add a prismatic command joint in metres per model, on a dummy child link, and re-point the existing mimics from the knuckle joint to it. The relation is exactly linear, so a plain <mimic multiplier="..." offset="..."/> covers it — offset is already used in robotiq_2f_140.xacro. Do not chain mimics; resolution is single-level in both robot_state_publisher and ros2_control.
3. Keep the radians joint for one release, behind an opt-out xacro flag, so consumers migrate on their own schedule. Humble's retirement on 2027-05-31 is a natural sunset. Both command interfaces start NaN and ros2_control claims command interfaces exclusively, so an unclaimed one is never written — "the non-NaN one wins" is a direct read of what the controller manager handed out, not a heuristic. Two claimed at once is a configuration error to report.
4. Derive, do not duplicate. The metres mapping comes from the SDK (units::openingFromRegister / openingToRegister); the radians one is derived from it, not a second independent mapping off gPO. Otherwise this reintroduces exactly the duplication Consume the SDK's SI conversions #70 exists to remove.

Knock-ons

The driver reads info_.joints.at(0) and is already joint-name agnostic, so the rename costs it nothing. What changes is the scale source: gripper_closed_position (one number) becomes a DeviceProfile. That parameter shape is shared with #44, #45 and #46, so it should be settled once.

Acceptance criterion: all six DeviceProfile fields come from the description, the count band included. The SDK's arithmetic is already parameterized — openingFromRegister / openingToRegister work off profile.registerPositionRange(), so a hardcoded band cannot exist inside the conversions — but profiles::k2F85 hardcodes openPosition 3 and closedPosition 230. A driver that passes the shipped profile has moved the hardcoding to another repo rather than removed it.

Those two counts are installation state, not model constants: non-standard fingertips are common and meet before the linkage runs out of travel, so the closed end lands below 230 (robotiq/grippers#34 has the detail and the bench test that confirms it). Today gripper_scaling.hpp hardcodes 3 and 230 for every model and takes gripper_closed_position as its only position parameter, so an installation whose fingers meet early gets compressed feedback and intermediate commands that overshoot — a half-closed command reaches for count 116 when half the real travel is nearer 101.

The pattern already exists in the driver: gripper_min_speed / gripper_max_speed arrive as <hardware> parameters rather than from k2F85, because the description is the authority on the installation. openPosition / closedPosition should arrive the same way, with the shipped profile as the fallback for anyone who has not measured. DeviceProfile is a plain struct with public members, so the driver can construct one from parameters and the SDK needs no change.

The failure mode is quiet: the band inherits k2F85 while the parameterization is designed around opening, speed and force, and nobody notices until a gripper with custom tips goes to the wrong place.

Numbers that are silently in the wrong unit once the command joint is metres:

Where	Today	In metres
goal_tolerance: 0.02 (both controller configs)	0.02 rad, about 2 mm	0.02 m is 20 mm, a quarter of the stroke
trigger_joint_command_threshold: 0.02 (topic-based sim)	0.02 rad	0.02 m is 23% of the stroke
max_velocity: 0.5 / max_effort: 50.0	rad/s	m/s
initial_value on the position state interface	${gripper_closed_position}	0.0
<limit velocity="0.5" effort="50"> per joint	rad/s	m/s
GRIPPER_JOINTS in robotiq_control.launch.py	joint names	already a map, just new names

Gazebo drops a link with no inertial, so the dummy link needs a token one or an xacro:unless sim_gazebo. The extra joint also appears in the MoveIt Setup Assistant's joint list and in /joint_states; a chain-based group or an SRDF that names joints explicitly will not pick it up, but it is worth a comment in the xacro and a line in the README.

The description tests are already parameterized on (model, joint), so a rename flows through them; test_ros2_control_urdf.py has a mimic-joint test to extend.

Blocks

#70 item 3, and with it the MCP's units.py / robot_description.py.

[bcastets-robotiq]

I would rather control the openijng in percentage to make it generic.

#95

[bcastets-robotiq]

Cross-reference: #95 proposes 0–1 ratios as the base command interface for position, speed and force (register = round(ratio × 255)), with the position in metres as an option that requires the gripper model (2F-85 or 2F-140).

This fits with this issue rather than replacing it:

• Joint states in metres, and the per-model opening conversion planned here, are what the metres option and /joint_states need.
• The SI command side changes role: the ratio is what users send by default, and metres is the opt-in for the position. Speed and force stay ratios in both modes.
• The direction differs: a larger opening is more open, while ratio 1.0 is fully closed.

See #95 for the details and the acceptance criteria.