Franka Panda and FR3¶

A Panda or FR3 accepts joint-space motion through the Franka Control Interface; the joint names and what libfranka keeps.

By the end of this page a Franka Emika Panda or Franka Research 3 accepts joint-space motion from Robot("panda", mode="real") through the Franka Control Interface, you know the joint names this package uses for each arm and why, and what is left to libfranka.

This needs an arm with FCI enabled in Desk, a workstation on the arm's network, and panda-py:

pip install panda-python          # the panda-py binding over libfranka (MIT); not a strands-robots extra

from strands_robots import Robot

arm = Robot("panda", mode="real", driver="strands", port="172.16.0.2")   # FrankaDriver; a bare call resolves to lerobot
print(arm.connect_eagerly())                                      # None, or why not
arm.send_action({f"joint{i}": 0.0 for i in range(1, 8)})         # all seven joints, radians

Joint names¶

FCI carries an unnamed seven-element vector. The names this driver uses are the names in each arm's own MuJoCo asset, so an action dict authored against Robot("panda") in simulation commands the real arm unchanged:

robot	joints	gripper key
`panda`	`joint1` ... `joint7`	`gripper_width`
`fr3`	`fr3_joint1` ... `fr3_joint7`	`gripper_width`
`fr3_v2`	`fr3v2_joint1` ... `fr3v2_joint7`	`gripper_width`

A joint-space command must name all seven joints with finite values; a partial dict or a key from the wrong arm is refused before libfranka sees it.

What the driver does¶

connect_eagerly() resolves panda_py, opens the FCI link and the Franka Hand. Off hardware it returns a reason string and leaves the driver usable: reads return their cache, writes refuse "not connected".
State: joint positions, velocities and link-side torques from one libfranka RobotState read, gripper width from the Hand. The arm sources state at 1 kHz (FCI_RATE_HZ); downsample_stride reports the ratio to the consumer's rate so a mesh publisher does not try to read every tick.
send_action hands the target to panda_py's guarded motion generator with speed_factor (default 0.2, a fifth of the arm's maximum), which owns the realtime loop and enforces the arm's limits.
stop() halts motion and reports.

What it does not do, on purpose¶

absent	why
a 1 kHz control loop, torque control	a Python thread that misses a tick triggers a reflex stop; libfranka's realtime context owns that loop
a joint-limit table	the Panda and the FR3 differ and this repo has neither to measure; libfranka refuses an out-of-envelope target and the refusal is reported verbatim
Cartesian control, kinematics	`O_T_EE` is on the state but not published as `_pose`, because the mesh's `_pose` is a base pose
`start_task`, `run_policy`	no policy provider emits Franka-shaped actions yet; both refuse and say so

Reading the arm from an agent¶

from strands import Agent

agent = Agent(tools=[arm])
agent("Read the joint positions and the measured torques. Do not move.")

The tool's read verbs are ungated. A motion verb pauses for approval (see agents).

Simulation¶

Robot("panda", keyframe="home"), Robot("fr3") and Robot("fr3_v2") build the MuJoCo twins from the menagerie assets. Because the joint names match, a policy or a recorded episode moves between the two modes without a remap.

Edit page