add kx2 backend

[rickwierenga]

No description provided.

[intrepid-pasha]

Hi @rickwierenga! First off thanks for the incredible work, we are big fans :).
Decided to test this PR on Hardware.
Pulling this branch down to try out for a KX-2 +
CS-10 plate-hotel setup, and we got it talking to a real robot today
on macOS 26.3 / Apple Silicon via PCBUSB (the MacCAN community
build that exposes PEAK's PCAN-Basic API on macOS). The interface="pcan"
default works as-is — no Mac-specific code changes needed.

Tested at HEAD dd6434b. Three small deltas to flag — the first two
look like things missed on the #1044 rebase pass (you already
re-parented KX2ArmBackend onto OrientableGripperArmBackend and
landed c1368ff, so most of that work is done); the third is an
over-cautious guard that's already obsolete given how the rest of the
stack is structured. Sharing in case any of these are useful — happy
to take them or leave them.

1. OrientableArm → OrientableGripperArm rename in kx2.py

KX2ArmBackend is now correctly subclassing OrientableGripperArmBackend
(👍), but pylabrobot/paa/kx2/kx2.py still imports the old OrientableArm
name and constructs it on self.arm. Every other consumer in the tree
already moved over, so I think this is just a missed file:

-from pylabrobot.capabilities.arms.orientable_arm import OrientableArm
+from pylabrobot.capabilities.arms.orientable_arm import OrientableGripperArm
@@
-    self.arm = OrientableArm(backend=backend, reference_resource=self.reference)
+    self.arm = OrientableGripperArm(backend=backend, reference_resource=self.reference)

2. KX2ArmBackend missing the new abstract members from
OrientableGripperArmBackend

OrientableGripperArmBackend is now abstract over move_gripper,
min_gripper_width, max_gripper_width — KX2ArmBackend extends
that base but doesn't implement those three, so today it's an abstract
class and KX2() can't instantiate. The existing open_gripper /
close_gripper on KX2ArmBackend (lines ~1207, ~1221) are still
referenced by the internal pick-up / drop sequences, so the minimal
shim is just to add the three new members and have move_gripper
dispatch into the existing two. Inserted right after is_gripper_closed:

@property
def min_gripper_width(self) -> Optional[float]:
    if self._config is None:
        return None
    axis_cfg = self._config.axes.get(Axis.SERVO_GRIPPER)
    return None if axis_cfg is None else axis_cfg.min_travel

@property
def max_gripper_width(self) -> Optional[float]:
    if self._config is None:
        return None
    axis_cfg = self._config.axes.get(Axis.SERVO_GRIPPER)
    return None if axis_cfg is None else axis_cfg.max_travel

async def move_gripper(
    self,
    width: float,
    force_sensing: bool = False,
    backend_params: Optional[BackendParams] = None,
) -> None:
    if force_sensing:
        await self.close_gripper(gripper_width=width, backend_params=backend_params)
    else:
        await self.open_gripper(gripper_width=width, backend_params=backend_params)

The properties intentionally read self._config (the Optional
form) instead of going through the _cfg accessor — that way they
correctly return None pre-setup(), which the abstract docstring
permits ("None if the gripper cannot be commanded …").

A note while you're in this area: the capability-level wrappers on
OrientableGripperArm itself (arm.open_gripper() / arm.close_gripper())
also lost their gripper_width= parameter — they now always drive to
max_/min_gripper_width. If you decide to push the same simplification
into KX2ArmBackend.open_gripper / close_gripper, we'd be happy to
adapt our test scripts; right now we route width-specific calls through
arm.move_gripper(width=…, force_sensing=…) which works fine.

3. has_rail=True is rejected up front, but the rest of the stack
already tolerates it

Our KX-2 has the rail axis on the bus (node 5). has_rail=False trips
node-list mismatch in KX2Driver.setup(); has_rail=True is hard-rejected
in KX2Driver.__init__. Reading the rest of the code, though:

• KX2ArmBackend._on_setup enables only MOTION_AXES = (SHOULDER, Z, ELBOW, WRIST) — the rail motor is never enabled.
• KX2ArmBackend._read_config already does its own rail detection from
  driver.node_id_list and warnings.warns "Rails has not been
  tested", then continues.
• request_joint_position hard-codes the four motion axes plus the
  servo gripper; rail position is silently dropped.
• kinematics.py doesn't reference KX2Config.robot_on_rail anywhere,
  so FK/IK/plan_joint_move are rail-blind.

So has_rail=True already means "rail on bus, dormant, ignored". The
guard in __init__ is the only thing in the way. Demoting it to a
warning let us run end-to-end on the real robot with the rail powered
up:

-    if has_rail or not has_servo_gripper:
-      raise NotImplementedError(
-        "KX2 has only been tested with the default 4-axis arm + servo "
-        "gripper topology (has_rail=False, has_servo_gripper=True). "
-        ...
-      )
+    if not has_servo_gripper:
+      raise NotImplementedError(
+        "KX2 has only been tested with the default servo-gripper "
+        "topology (has_servo_gripper=True). The gripper-less path has "
+        "shim code but the setup / homing layer needs work — see "
+        "KX2ArmBackend._on_setup and servo_gripper_initialize."
+      )
+    if has_rail:
+      warnings.warn(
+        "has_rail=True: rail (node 5) is included in CAN bus discovery "
+        "and PDO remapping, but the rail motor is NOT enabled and "
+        "kinematics / motion code paths ignore it. Rail position is "
+        "not queried. Verify your task does not need to move the rail."
+      )

(import warnings added near the other stdlib imports.) All 148 KX-2
tests still pass with this. Happy to flip back to NotImplementedError
if you prefer to keep the rail topology gated until full rail support
lands — totally fair call.

What's working on real hardware as of today

• Bus brought up over PCAN-USB → PCBUSB → python-can → canopen on
  macOS 26.3 / arm64. interface="pcan" default needed no changes.
• All six drives discovered ([1, 2, 3, 4, 5, 6]), motion axes enabled,
  rail dormant.
• servo_gripper_initialize ran the full hard-stop + index-pulse home
  routine successfully on the first try.
• request_gripper_pose, request_joint_position, motion_limits,
  get_estop_state all returning sensible values from the drives.
• arm.move_gripper(width=…, force_sensing=…) reaches commanded width.
• First closed-loop arm-axis motion: Z bump +30 mm at 30 mm/s.
  Arrived within 0.025 mm of target (tol 0.5 mm). 🎉

Multi-axis moves and rail-axis support are next on our side. Happy to
open a PR against your branch with these three patches if that's
useful, otherwise feel free to fold them in directly — really no
preference. Thanks again for the work on this branch; the
CANopen-direct approach is exactly what we needed, and it's
surprisingly close to working out of the box.