franka-rs for Python

import franka: drive a Franka FR3 (FCI v10) or Franka Emika Robot / Panda (FCI v5) from a program that runs at its own rate -- a policy at 5 to 30 Hz, a script, a keyboard. The 1 kHz control loop is franka-rs's target control on a Rust thread; the GIL is never on the 1 kHz path. move_to / move_by are one write into that loop's seqlock slot, follow hands a chunk of targets to a Rust timer thread, and state() is a copy out of a mutex.

For an arm served by franka-node over Zenoh, see its client, pip install franka-node-client.

import franka

robot = franka.Robot("172.16.0.2")    # FRANKA_REALTIME=ignore for franka-sim

with robot.cartesian_targets(max_velocity=0.3, max_acceleration=0.5, max_jerk=20.0) as arm:
    while not done:
        obs = arm.state().flat()      # 1-D float64 in the order of franka.FLAT_LAYOUT
        action = policy(obs)
        arm.move_by(action[:3])       # metres, relative to the current target
        # or: arm.move_to([x, y, z]); arm.follow(chunk, dt=0.02)  # chunk: (N, 3) absolute

robot.joint_targets(fraction=0.2) is the same over the seven joints. __exit__ stops the loop; an exception inside the block still stops the motion and is re-raised. Errors are franka.FrankaError; a motion the robot aborted is a franka.ControlException with the reflex reason. robot.gripper() drives the Franka Hand, releasing the GIL while it waits.

The loop tracks the targets with the crate's own impedance torques by default, so the arm is a spring around the target: push it and it gives way, let go and it returns. cartesian_stiffness=400 softens the default 750 N/m (15 Nm/rad rotational); cartesian_damping, joint_stiffness, joint_damping, torque_limits, posture, leash and project_joint_gains are the other knobs, and backend='robot' has the robot's own controller track the targets instead, as the bindings did in 0.2.0. Target control sets no collision thresholds: with the default gains give robot.set_collision_behavior_simple(...) at least 40 N / 40 Nm. The law and the defaults are on the book's Python page.

A Cartesian target of 7 elements (move_to, the rows of follow) sets the orientation too, as a unit quaternion x, y, z, w after the position; a move_by of 6 elements appends a rotation vector (axis times angle, rad) composed onto the target orientation in the base frame. arm.target() is that 7-vector, arm.target_pose() the 4x4 matrix. examples/policy_loop.py is a jittery 6-10 Hz loop with a circle, a yaw and tilt sweep and a follow chunk; examples/rotate.py is the rotation alone, and examples/quickstart.ipynb a notebook that connects, moves, and replays the motion inline with Franka's meshes. robot.model() is the kinematics and dynamics over numpy (pose, body_jacobian, zero_jacobian, mass, coriolis, gravity, link_poses, hand_pose).

Install

pip install franka-rs

The wheel is named franka-rs, the package franka. To build from the source tree into a virtualenv instead (prefix with env -u CONDA_PREFIX under an active conda env):

pip install maturin
maturin develop --release -m crates/franka-py/Cargo.toml

The crate is publish = false on crates.io; its tests are

flock .sim.lock env FRANKA_SIM_IMAGE=franka-sim:dev pytest crates/franka-py/tests

against the simulator (see the book's Testing with franka-sim).