Franky

High-Level Control Library for Franka Robots with Python and C++ Support

Franky is a high-level control library for Franka robots offering Python and C++ support. By wrapping libfranka in a Python interface, Franky eliminates the need for strict real-time programming at 1 kHz. Instead, you can define higher-level motion targets in Python, and Franky will use Ruckig to plan time-optimal trajectories in real time.

Although Python does not provide real-time guarantees, Franky strives to maintain as much real-time control as possible. Motions can be preempted at any moment, prompting Franky to replan trajectories on the fly. To handle unforeseen situations—such as unexpected contact with the environment—Franky includes a reaction system that allows to dynamically update motion commands.

Check out the tutorial and the examples for an introduction. The full documentation can be found at https://timschneider42.github.io/franky/.

Differences to frankx

Franky is a fork of frankx, though both codebase and functionality differ substantially from frankx by now. In particular, franky provides the following new features/improvements:

• Motions can be updated asynchronously.
• Reactions allow for the registration of callbacks instead of just printing to stdout when fired.
• The robot state is also available during control.
• A larger part of the libfranka API is exposed to python (e.g.,setCollisionBehavior, setJoinImpedance, and setCartesianImpedance).
• Cartesian motion generation handles boundaries in Euler angles properly.
• There is a new joint motion type that supports waypoints.
• The signature of `Affine` changed. Affine does not handle elbow positions anymore. Instead, a new class RobotPose stores both the end-effector pose and optionally the elbow position.
• The MotionData class does not exist anymore. Instead, reactions and other settings moved to Motion.
• The `Measure` class allows for arithmetic operations.
• Exceptions caused by libfranka are raised properly instead of being printed to stdout.
• We provide wheels for both Franka Research 3 and the older Franka Panda

Setup

To install franky, you have to follow three steps:

1. Ensure that you are using a realtime kernel
2. Ensure that the executing user has permission to run real-time applications
3. Install franky via pip or build it from the sources

Installing a real-time kernel

In order for franky to function properly, it requires the underlying OS to use a realtime kernel. Otherwise, you might see communication_constrains_violation errors.

To check whether your system is currently using a real-time kernel, type uname -a. You should see something like this:

$ uname -a
Linux [PCNAME] 5.15.0-1056-realtime #63-Ubuntu SMP PREEMPT_RT ...

If it does not say PREEMPT_RT, you are not currently running a real-time kernel.

There are multiple ways of installing a real-time kernel. You can build it from source or, if you are using Ubuntu, it can be enabled through Ubuntu Pro.

Allowing the executing user to run real-time applications

First, create a group realtime and add your user (or whoever is running franky) to this group:

sudo addgroup realtime
sudo usermod -a -G realtime $(whoami)

Afterward, add the following limits to the real-time group in /etc/security/limits.conf:

@realtime soft rtprio 99
@realtime soft priority 99
@realtime soft memlock 102400
@realtime hard rtprio 99
@realtime hard priority 99
@realtime hard memlock 102400

Log out and log in again to let the changes take effect.

To verify that the changes were applied, check if your user is in the realtime group:

$ groups
... realtime

If realtime is not listed in your groups, try rebooting.

Installing franky

To start using franky with Python and libfranka 0.15.0, just install it via

pip install franky-panda

We also provide wheels for libfranka versions 0.7.1, 0.8.0, 0.9.2, 0.10.0, 0.11.0, 0.12.1, 0.13.3, 0.14.2, and 0.15.0. They can be installed via

VERSION=0-9-2
wget https://github.com/TimSchneider42/franky/releases/latest/download/libfranka_${VERSION}_wheels.zip
unzip libfranka_${VERSION}_wheels.zip
pip install numpy
pip install --no-index --find-links=./dist franky-panda

Franky is based on libfranka, Eigen for transformation calculations and pybind11 for the Python bindings. As the Franka is sensitive to acceleration discontinuities, it requires jerk-constrained motion generation, for which franky uses the Ruckig community version for Online Trajectory Generation (OTG).

After installing the dependencies (the exact versions can be found here), you can build and install franky via

git clone --recurse-submodules git@github.com:timschneider42/franky.git
cd franky
mkdir -p build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make
make install

To use franky, you can also include it as a subproject in your parent CMake via add_subdirectory(franky) and then target_link_libraries(<target> franky).

If you need only the Python module, you can install franky via

pip install .

Make sure that the built library _franky.cpython-3**-****-linux-gnu.so is in the Python path, e.g. by adjusting PYTHONPATH accordingly.

Using Docker

To use franky within Docker we provide a Dockerfile and accompanying docker-compose file.

git clone https://github.com/timschneider42/franky.git
cd franky/
docker compose build franky-run

To use another version of libfranka than the default (0.15.0) add a build argument:

docker compose build franky-run --build-arg LIBFRANKA_VERSION=0.9.2

To run the container:

docker compose run franky-run bash

The container requires access to the host machines network and elevated user rights to allow the docker user to set RT capabilities of the processes run from within it.

Building franky with Docker

For building franky and its wheels, we provide another Docker container that can also be launched using docker-compose:

docker compose build franky-build
docker compose run --rm franky-build run-tests  # To run the tests
docker compose run --rm franky-build build-wheels  # To build wheels for all supported python versions

Tutorial

Franky comes with both a C++ and Python API that differ only regarding real-time capability. We will introduce both languages next to each other. In your C++ project, just include include <franky.hpp> and link the library. For Python, just import franky. As a first example, only four lines of code are needed for simple robotic motions.

 ++
#include <franky.hpp>
using namespace franky;
 
// Connect to the robot with the FCI IP address
Robot robot("172.16.0.2");
 
// Reduce velocity and acceleration of the robot
robot.setRelativeDynamicsFactor(0.05);
 
// Move the end-effector 20cm in positive x-direction
auto motion = std::make_shared<CartesianMotion>(RobotPose(Affine({0.2, 0.0, 0.0}), 0.0), ReferenceType::Relative);
 
// Finally move the robot
robot.move(motion);

The corresponding program in Python is

from franky import Affine, CartesianMotion, Robot, ReferenceType
 
robot = Robot("172.16.0.2")
robot.relative_dynamics_factor = 0.05
 
motion = CartesianMotion(Affine([0.2, 0.0, 0.0]), ReferenceType.Relative)
robot.move(motion)

Furthermore, we will introduce methods for geometric calculations, for moving the robot according to different motion types, how to implement real-time reactions and changing waypoints in real time as well as controlling the gripper.

Geometry

franky.Affine is a python wrapper for Eigen::Affine3d. It is used for Cartesian poses, frames and transformation. franky adds its own constructor, which takes a position and a quaternion as inputs:

import math
from scipy.spatial.transform import Rotation
from franky import Affine
 
z_translation = Affine([0.0, 0.0, 0.5])
 
quat = Rotation.from_euler("xyz", [0, 0, math.pi / 2]).as_quat()
z_rotation = Affine([0.0, 0.0, 0.0], quat)
 
combined_transformation = z_translation * z_rotation

In all cases, distances are in [m] and rotations in [rad].

Robot

We wrapped most of the libfanka API (including the RobotState or ErrorMessage) for Python. Moreover, we added methods to adapt the dynamics of the robot for all motions. The rel name denotes that this a factor of the maximum constraints of the robot.

from franky import Robot
 
robot = Robot("172.16.0.2")
 
# Recover from errors
robot.recover_from_errors()
 
# Set velocity, acceleration and jerk to 5% of the maximum
robot.relative_dynamics_factor = 0.05
 
# Alternatively, you can define each constraint individually
robot.velocity_rel = 0.2
robot.acceleration_rel = 0.1
robot.jerk_rel = 0.01
 
# Get the current pose
current_pose = robot.current_pose

Robot State

The robot state can be retrieved by calling the following methods:

• state: Object of type franky.RobotState, which extends the libfranka franka::RobotState structure by additional state elements.
• current_cartesian_state: Object of type franky.CartesianState, which contains the end-effector pose and velocity obtained from franka::RobotState::O_T_EE and franka::RobotState::O_dP_EE_c.
• current_joint_position: Object of type franky.JointState, which contains the joint positions and velocities obtained from franka::RobotState::q and franka::RobotState::dq.

robot = Robot("172.16.0.2")
 
# Get the current state as raw `franky.RobotState`
state = robot.state
 
# Get the robot's cartesian state
cartesian_state = robot.current_cartesian_state
robot_pose = cartesian_state.pose  # Contains end-effector pose and elbow position
ee_pose = robot_pose.end_effector_pose
elbow_pos = robot_pose.elbow_position
robot_velocity = cartesian_state.velocity  # Contains end-effector twist and elbow velocity
ee_twist = robot_velocity.end_effector_twist
elbow_vel = robot_velocity.elbow_velocity
 
# Get the robot's joint state
joint_state = robot.current_joint_state
joint_pos = joint_state.position
joint_vel = joint_state.velocity

Motion Types

Franky defines a number of different motion types. In python, you can use them as follows:

import math
from scipy.spatial.transform import Rotation
from franky import JointWaypointMotion, JointWaypoint, JointPositionStopMotion, CartesianMotion, \
    CartesianWaypointMotion, CartesianWaypoint, Affine, Twist, RobotPose, ReferenceType, CartesianPoseStopMotion, \
    CartesianState, JointState
 
# A point-to-point motion in the joint space
m1 = JointWaypointMotion([JointWaypoint([-0.3, 0.1, 0.3, -1.4, 0.1, 1.8, 0.7])])
 
# A motion in joint space with multiple waypoints
m2 = JointWaypointMotion([
    JointWaypoint([-0.3, 0.1, 0.3, -1.4, 0.1, 1.8, 0.7]),
    JointWaypoint([0.0, 0.3, 0.3, -1.5, -0.2, 1.5, 0.8]),
    JointWaypoint([0.1, 0.4, 0.3, -1.4, -0.3, 1.7, 0.9])
])
 
# Intermediate waypoints also permit to specify target velocities. The default target velocity is 0, meaning that the
# robot will stop at every waypoint.
m3 = JointWaypointMotion([
    JointWaypoint([-0.3, 0.1, 0.3, -1.4, 0.1, 1.8, 0.7]),
    JointWaypoint(
        JointState(
            position=[0.0, 0.3, 0.3, -1.5, -0.2, 1.5, 0.8],
            velocity=[0.1, 0.0, 0.0, 0.0, -0.0, 0.0, 0.0])),
    JointWaypoint([0.1, 0.4, 0.3, -1.4, -0.3, 1.7, 0.9])
])
 
# Stop the robot
m4 = JointPositionStopMotion()
 
# A linear motion in cartesian space
quat = Rotation.from_euler("xyz", [0, 0, math.pi / 2]).as_quat()
m5 = CartesianMotion(Affine([0.4, -0.2, 0.3], quat))
m6 = CartesianMotion(RobotPose(Affine([0.4, -0.2, 0.3], quat), elbow_position=0.3))  # With target elbow angle
 
# A linear motion in cartesian space relative to the initial position
# (Note that this motion is relative both in position and orientation. Hence, when the robot's end-effector is oriented
# differently, it will move in a different direction)
m7 = CartesianMotion(Affine([0.2, 0.0, 0.0]), ReferenceType.Relative)
 
# Generalization of CartesianMotion that allows for multiple waypoints
m8 = CartesianWaypointMotion([
    CartesianWaypoint(RobotPose(Affine([0.4, -0.2, 0.3], quat), elbow_position=0.3)),
    # The following waypoint is relative to the prior one and 50% slower
    CartesianWaypoint(Affine([0.2, 0.0, 0.0]), ReferenceType.Relative, RelativeDynamicsFactor(0.5, 1.0, 1.0))
])
 
# Cartesian waypoints also permit to specify target velocities
m9 = CartesianWaypointMotion([
    CartesianWaypoint(Affine([0.5, -0.2, 0.3], quat)),
    CartesianWaypoint(
        CartesianState(
            pose=Affine([0.4, -0.1, 0.3], quat),
            velocity=Twist([-0.01, 0.01, 0.0]))),
    CartesianWaypoint(Affine([0.3, 0.0, 0.3], quat))
])
 
# Stop the robot. The difference of JointPositionStopMotion to CartesianPoseStopMotion is that JointPositionStopMotion
# stops the robot in joint position control mode while CartesianPoseStopMotion stops it in cartesian pose control mode.
# The difference becomes relevant when asynchronous move commands are being sent (see below).
m10 = CartesianPoseStopMotion()

Every motion and waypoint type allows to adapt the dynamics (velocity, acceleration and jerk) by setting the respective relative_dynamics_factor parameter.

The real robot can be moved by applying a motion to the robot using move:

robot.move(m1)
robot.move(m2)

Real-Time Reactions

By adding reactions to the motion data, the robot can react to unforeseen events. In the Python API, you can define conditions by using a comparison between a robot's value and a given threshold. If the threshold is exceeded, the reaction fires.

from franky import CartesianMotion, Affine, ReferenceType, Measure, Reaction
 
motion = CartesianMotion(Affine([0.0, 0.0, 0.1]), ReferenceType.Relative)  # Move down 10cm
 
reaction_motion = CartesianMotion(Affine([0.0, 0.0, 0.01]), ReferenceType.Relative)  # Move up for 1cm
 
# Trigger reaction if the Z force is greater than 30N
reaction = Reaction(Measure.FORCE_Z > 30.0, reaction_motion)
motion.add_reaction(reaction)
 
robot.move(motion)

Possible values to measure are

• Measure.FORCE_X, Measure.FORCE_Y, Measure.FORCE_Z: Force in X, Y and Z direction
• Measure.REL_TIME: Time in seconds since the current motion started
• Measure.ABS_TIME: Time in seconds since the initial motion started

The difference between Measure.REL_TIME and Measure.ABS_TIME is that Measure.REL_TIME is reset to zero whenever a new motion starts (either by calling Robot.move or as a result of a triggered Reaction). Measure.ABS_TIME, on the other hand, is only reset to zero when a motion terminates regularly without being interrupted and the robot stops moving. Hence, Measure.ABS_TIME measures the total time in which the robot has moved without interruption.

Measure values support all classical arithmetic operations, like addition, subtraction, multiplication, division, and exponentiation (both as base and exponent).

normal_force = (Measure.FORCE_X ** 2 + Measure.FORCE_Y ** 2 + Measure.FORCE_Z ** 2) ** 0.5

With arithmetic comparisons, conditions can be generated.

normal_force_within_bounds = normal_force < 30.0
time_up = Measure.ABS_TIME > 10.0

Conditions support negation, conjunction (and), and disjunction (or):

abort = ~normal_force_within_bounds | time_up
fast_abort = ~normal_force_within_bounds | time_up

To check whether a reaction has fired, a callback can be attached:

from franky import RobotState
 
def reaction_callback(robot_state: RobotState, rel_time: float, abs_time: float):
    print(f"Reaction fired at {abs_time}.")
 
reaction.register_callback(reaction_callback)

Note that these callbacks are not executed in the control thread since they would otherwise block it. Instead, they are put in a queue and executed by another thread. While this scheme ensures that the control thread can always run, it cannot prevent that the queue grows indefinitely when the callbacks take more time to execute than it takes for new callbacks to be queued. Hence, callbacks might be executed significantly after their respective reaction has fired if they are triggered in rapid succession or take a long time to execute.

In C++ you can additionally use lambdas to define more complex behaviours:

 ++
auto motion = CartesianMotion(RobotPose(Affine({0.0, 0.0, 0.2}), 0.0), ReferenceType::Relative);
 
// Stop motion if force is over 10N
auto stop_motion = StopMotion<franka::CartesianPose>()
 
motion
  .addReaction(
    Reaction(
      Measure::ForceZ() > 10.0,  // [N],
      stop_motion))
  .addReaction(
    Reaction(
      Condition(
        [](const franka::RobotState& state, double rel_time, double abs_time) {
          // Lambda condition
          return state.current_errors.self_collision_avoidance_violation;
        }),
      [](const franka::RobotState& state, double rel_time, double abs_time) {
        // Lambda reaction motion generator
        // (we are just returning a stop motion, but there could be arbitrary 
        // logic here for generating reaction motions)
        return StopMotion<franka::CartesianPose>();
      })
    ));
 
robot.move(motion)

Real-Time Motion

By setting the asynchronous parameter of Robot.move to True, the function does not block until the motion finishes. Instead, it returns immediately and, thus, allows the main thread to set new motions asynchronously.

import time
from franky import Affine, CartesianMotion, Robot, ReferenceType
 
robot = Robot("172.16.0.2")
robot.relative_dynamics_factor = 0.05
 
motion1 = CartesianMotion(Affine([0.2, 0.0, 0.0]), ReferenceType.Relative)
robot.move(motion1, asynchronous=True)
 
time.sleep(0.5)
motion2 = CartesianMotion(Affine([0.2, 0.0, 0.0]), ReferenceType.Relative)
robot.move(motion2, asynchronous=True)

By calling Robot.join_motion the main thread can be synchronized with the motion thread, as it will block until the robot finishes its motion.

robot.join_motion()

Note that when exceptions occur during the asynchronous execution of a motion, they will not be thrown immediately. Instead, the control thread stores the exception and terminates. The next time Robot.join_motion or Robot.move are called, they will throw the stored exception in the main thread. Hence, after an asynchronous motion has finished, make sure to call Robot.join_motion to ensure being notified of any exceptions that occurred during the motion.

Gripper

In the franky::Gripper class, the default gripper force and gripper speed can be set. Then, additionally to the libfranka commands, the following helper methods can be used:

 ++
#include <franky.hpp>
#include <chrono>
#include <future>
 
auto gripper = franky::Gripper("172.16.0.2");
 
double speed = 0.02; // [m/s]
double force = 20.0; // [N]
 
// Move the fingers to a specific width (5cm)
bool success = gripper.move(0.05, speed);
 
// Grasp an object of unknown width
success &= gripper.grasp(0.0, speed, force, epsilon_outer=1.0);
 
// Get the width of the grasped object
double width = gripper.width();
 
// Release the object
gripper.open(speed);
 
// There are also asynchronous versions of the methods
std::future<bool> success_future = gripper.moveAsync(0.05, speed);
 
// Wait for 1s
if (!success_future.wait_for(std::chrono::seconds(1)) == std::future_status::ready) {
  // Get the result
  std::cout << "Success: " << success_future.get() << std::endl;
} else {
  gripper.stop();
  success_future.wait();
  std::cout << "Gripper motion timed out." << std::endl;
}

The Python API follows the c++ API closely:

import franky
 
gripper = franky.Gripper("172.16.0.2")
 
speed = 0.02  # [m/s]
force = 20.0  # [N]
 
# Move the fingers to a specific width (5cm)
success = gripper.move(0.05, speed)
 
# Grasp an object of unknown width
success &= gripper.grasp(0.0, speed, force, epsilon_outer=1.0)
 
# Get the width of the grasped object
width = gripper.width
 
# Release the object
gripper.open(speed)
 
# There are also asynchronous versions of the methods
success_future = gripper.move_async(0.05, speed)
 
# Wait for 1s
if success_future.wait(1):
    print(f"Success: {success_future.get()}")
else:
    gripper.stop()
    success_future.wait()
    print("Gripper motion timed out.")

Development

Franky is written in C++17 and Python3.7. It is currently tested against following versions

• Libfranka 0.7.1, 0.8.0, 0.9.2, 0.10.0, 0.11.0, 0.12.1, 0.13.3, 0.14.2, 0.15.0
• Eigen 3.4.0
• Pybind11 2.13.6
• Pinocchio 3.4.0
• Python 3.7, 3.8, 3.9, 3.10, 3.11, 3.12, 3.13
• Catch2 2.13.8 (for testing only)

License

For non-commercial applications, this software is licensed under the LGPL v3.0. If you want to use franky within commercial applications or under a different license, please contact us for individual agreements.