Skip to content

Installation & connection

Run the C++ daemon on the NUC connected to your FR3, and install the Python client on your workstation. Use the NUC address when connecting from Python.

NUC side

Ships as a Docker image (libfranka 0.17.0, pinocchio 3.x, Eigen 3.4, gcc 13).

Host kernel must be PREEMPT_RT

Build & run

./fr3-stack build              # one time (~5–10 min, cached after)
./fr3-stack up cart            # cartesian impedance, anchored at current pose
./fr3-stack up hybrid --ft     # hybrid force/position with Bota FT sensor
./fr3-stack up idle            # gravity-comp + damping (hand-guidable)
./fr3-stack up cart --attach   # foreground (default is detached)
./fr3-stack logs -f
./fr3-stack down

Modes are idle, cart, joint, adm, and hybrid. For a non-idle initial controller, the daemon anchors its target to the live robot state at activation.

Daemon flags

Flag Default Notes
--robot <ip> required FR3 IP
--cmd-port 5555 workstation→NUC commands
--state-port 5556 NUC→workstation state (~200 Hz)
--initial-controller idle idle | cartesian_impedance | hybrid
--ft-sensor-kind empty currently only bota
--ft-sensor-config empty Bota: path to driver JSON

Shorthands: --idle, --cart, --hybrid.

Compose env vars

FR3_ROBOT_IP=192.168.1.11 \
FR3_INITIAL_CONTROLLER=cartesian_impedance \
FR3_FT_SENSOR_KIND=bota \
FR3_FT_SENSOR_CONFIG=/opt/bota/driver_config/bota_binary.json \
./fr3-stack up cart --ft

F/T sensor (optional)

The Bota driver is bundled under third_party/. Select the FT backend on the NUC:

./fr3-stack up hybrid --ft

When publishing, State.wrench_ft is a length-6 vector; otherwise None. send_hybrid refuses to start without one unless require_ft_sensor=False.

Workstation side

pip install -e .

Connection config (fr3.yaml)

Read from one of:

  1. $FR3_CONFIG (explicit override)
  2. ./fr3.yaml (per-project, gitignored)
  3. ~/.config/fr3/config.yaml (user-global)

Template (examples/fr3.example.yaml):

nuc_host:  192.168.1.8       # where fr3-stack listens
robot_ip:  192.168.1.11      # FR3's own IP
desk:
  user:     YOUR_DESK_USER
  password: YOUR_DESK_PASSWORD

Env vars override file values: FR3_NUC_HOST, FR3_ROBOT_IP, FR3_DESK_USER, FR3_DESK_PASS.

Two API layers

Class Module Use when
Robot fr3_stack.robot Streaming wire — one send_* per command. Direct access to every wire field.
Arm fr3_stack.client Pose-centric facade for inference / teleop loops. arm.robot is the escape hatch.

Hello world (Robot)

from fr3_stack import Robot
import numpy as np

with Robot("nuc.local") as robot:
    s = robot.wait_for_state()
    robot.send_cartesian_impedance(
        target_pos       = s.pos + np.array([0, 0, 0.05]),
        target_quat_xyzw = s.quat_xyzw,
        K                = [200, 200, 800, 30, 30, 30],
    )
    print(robot.state.pos, robot.state.wrench_ext)

Per-call kwargs (K=, D=, target_wrench=, …) update an internal last-sent cache, so subsequent calls without those kwargs reuse the new values.

Pose-centric API (Arm)

Stiffness is sticky — set once, then stream Pose targets:

from fr3_stack import Arm, Pose

with Arm("nuc.local") as arm:
    arm.set_stiffness(K=[200, 200, 800, 30, 30, 30], damp_ratio=0.9)
    arm.move_to(Pose([0.5, 0.0, 0.4], [0, 0, 0, 1]), duration=2.0)

    obs    = arm.observe()
    target = Pose(obs.pose.pos + [0, 0, 0.05], obs.pose.quat)
    arm.send(target)

    arm.hold()        # lock at current pose with cached gains
    # arm.relax()     # gravity-comp + damping (hand-guidable; sends idle)

arm.move_to(target, duration=T) blocks for T seconds while the daemon's min-jerk generator drives the move; afterwards the controller holds at the goal. arm.send(...) is fire-and-forget streaming. Hybrid commands drop through to arm.robot.send_hybrid(...).

Per-controller defaults

K/D/M defaults live in fr3_stack/configs/<name>.yaml, loaded when Robot is constructed. Profile variants (<name>.<profile>.yaml) let you keep multiple presets:

robot = Robot("nuc.local", profiles={"cartesian_impedance": "stiff"})
robot.set_profile("hybrid", "polish")        # swap at runtime

F/T calibration

If you have a sensor mounted past the wrist, run payload identification before using hybrid:

./fr3-stack up hybrid --ft

# manual: hand-guide to ≥6 static poses with varied EE tilt
fr3-ft-calibrate 192.168.1.8

# or record once, replay automatically (CLEAR the workspace)
fr3-ft-calibrate-record 192.168.1.8
fr3-ft-calibrate-auto   192.168.1.8

The solver writes fr3_stack/sensors/bota/config/ft_calibration.yaml, which docker-compose bind-mounts into the daemon. The daemon reads it once at startup — restart to pick up a new file:

./fr3-stack down
./fr3-stack up hybrid --ft

After that, state.wrench_ft is compensated, state.wrench_ft_raw carries the raw stream, and state.ft_compensated == True.

Termination

robot.terminate()    # sends idle with termination=True; daemon stops RT loop