Build & Setup Guide · Updated Nov 2026
A complete, reproducible build of a GELLO teleoperation arm for the Franka Research 3: a $323 bill of materials, 3D-printing and assembly notes, Dynamixel wiring and ID setup, calibration against the robot, and the mistakes that cost the most time. The same arm went on to collect our lab's 10-task FR3 demonstration dataset.
The finished leader arm driving the FR3 in real time — each run recorded as a demonstration for imitation learning.
At a glance
Scope. The hardware build plus the software needed to verify it. Franka FCI setup, the controller stack, and policy training live on the FR3 cell page.
01 · Introduction

GELLO is a low-cost, 3D-printed leader arm: a passive kinematic twin of the robot that you move by hand while the real arm follows, so demonstrations come out smooth and intuitive, in the robot's own joint space, without a VR headset or a SpaceMouse.
This guide documents the exact build used on our lab's Franka FR3: what to buy, how to assemble and wire it, how to calibrate it against the robot, and how to confirm it works before the robot ever moves. The same leader arm has since collected the lab's 10-task FR3 demonstration dataset (§8).
gello_mechanical/franka_fr3 design./dev/gello by a udev rule.gello_software (Dynamixel driver + offset calibration) and RLinf's real-world environment: GelloJointIntervention turns leader joints into joint-velocity commands for the Polymetis / DROID controller on the NUC.02 · Bill of Materials
Everything I actually used — exact models, quantities, and where I bought them. The seven joint servos are the bulk of the cost; the structural parts were printed in-house.
| Item | Spec / model | Qty | Unit | Source |
|---|---|---|---|---|
| StructuralFR3 GELLO 3D-printed partsPrinted on a Bambu. Tolerances matter — some holes need sanding, drilling, or heat-set tweaks. | 18 pieces · gello_mechanical/franka_fr3 | 18 | in-house | GitHub ↗ |
| ActuatorArm joint motorsMain joint servos. Use the TTL 3-pin version — not the RS-485 4-pin. | DYNAMIXEL XL330-M288-T | 7 | $27.49 | ROBOTIS ↗ |
| Frame / IdlerHinge frame + idler setXL330 hinge/idler side. Keep spares — small plastic parts are easy to lose. | FPX330-H101 (4 pcs) | 1 set | $9.80 | ROBOTIS ↗ |
| ActuatorTrigger / gripper motorGripper trigger input. Needs its own gripper-range calibration in software. | DYNAMIXEL XL330-M077-T | 1 | $27.49 | ROBOTIS ↗ |
| ControllerGELLO controllerMotor cable into a white Dynamixel port; USB-C to the PC. | OpenRB-150 | 1 | $28.64 | ROBOTIS ↗ |
| CableLong Dynamixel cableFor longer joint routing. Must be the 3-pin TTL JST-JST cable. | Robot Cable-X3P JST-JST, 360 mm | 3 | $10.95 | MyBotShop (ROBOTIS reseller) ↗ |
| CableUSB cable for OpenRBMust support data — charge-only cables power the board but hide the serial device. | USB-A → USB-C, ~2 m | 1 | $8.00 | Amazon (search) ↗ |
| PowerExternal DXL power supplyOptional for passive reading; required for torque / damping / virtual-spring mode. Never 12V on XL330. | 5V, ≥3A (5A preferred) | 1 | $12.00 | Amazon (search) ↗ |
| ToolScrewsKeep several lengths — some printed holes are shallow. | M2 & M3 self-tapping | 1 set | $12.00 | Amazon (search) ↗ |
| SpringJoint springs (J2 & J4)Both from one stainless torsion-spring assortment kit (box label 1.2*10*35mm-180°). The kit lets you test-fit stiffnesses. | Torsion, 1.2 × 10 × 35 mm, 180° | 2 | $11.40 / kit | Amazon (search) ↗ |
| Total | ≈ $323 | $335 with the 5V supply · prints in-house | ||
03 · Mechanical Assembly

The arm is 18 printed pieces from the franka_fr3 design. It doesn't carry the load a real FR3 does, but the joints still take repeated screw pressure and hand force during teleop — so I printed at ~50% infill, favoring strength over a featherweight part.

Two torsion springs (the same 1.2 × 10 × 35 mm, 180° size) sit at joints 2 and 4 to offset gravity so the passive arm floats instead of sagging. The mount angle and pretension are the fiddly part: too stiff and the arm fights your hand, too loose and it droops. I set them so the arm holds a neutral pose on its own with almost no effort to move.
Route each lead along its link with enough slack to survive the joint's full range — a cable that's taut at one extreme will unseat its connector or cap the motion. Finally, fix the base so the leader can't tip during teleop and its reach maps comfortably onto the robot's workspace.
04 · Wiring
The XL330s share a single 3-wire TTL bus: each servo has two ports, so you chain them servo-to-servo from the base out to the gripper. The first servo plugs into the OpenRB-150's white Dynamixel port; the board is the bus master and talks to the PC over a single USB-C cable that carries both power and data.
It enumerates as a USB CDC-ACM serial device — on my machine /dev/ttyACM0(note: ACM, not the FTDI ttyUSB* some docs assume). Confirm it before going further:
ls -l /dev/serial/by-id/ # stable name for the OpenRB-150 ls -l /dev/ttyACM0 # what gello_software expects
This is the step that bites everyone: every Dynamixel ships with the same factory ID (1). Put them all on one bus and they answer at once and collide, so a scan flickers between seeing one device and none. Before building the daisy-chain, connect one motor at a time and use ROBOTIS's software (Dynamixel Wizard 2.0, through the OpenRB-150) to write a unique ID to each (default baud 57600) — 1→7 for the joints from base to tip, and 8 for the gripper. Only one servo may be on the bus while you write, or the new ID gets broadcast to every ID-1 servo at once.

The stock Dynamixel leads didn't reach across the longer links of this build, so I made my own: cut to length, soldered, and heat-shrunk each joint's extension. It's tedious but worth doing carefully — a cold joint here turns into the kind of intermittent bus fault that's miserable to chase later.
For passive teleoperation — how GELLO is used here, with servo torque off — the OpenRB-150's USB-C power is enough; the arm only reads joint angles. An external supply (5V, ≥3A — 5A preferred, never 12V on XL330) only matters if you actually drive the servos: torque, damping, or virtual-spring mode. I first thought it was required and it wasn't — a misread of a torque-enable warning.
05 · Common Pitfalls
Pitch offsets look right but are 90° off
Joints 2, 4 and 6 cannot be calibrated by holding the leader in free space: the arm looks matched while an offset is a quarter turn out. Calibrate against the homed robot, a rigid reference, and validate each joint's sign and offset against its known pose.
Calibration lost with the container
Editing the installed gello_software config inside a container is lost on rebuild — ours dropped the June calibration in July. Keep the leader's DynamixelRobotConfig in your own repo and inject it into PORT_CONFIG_MAP at runtime.
GELLO ↔ FR3 pose mismatch
Calibration assumes leader and follower start matched. If they disagree at startup, every command is offset — and the leader can sit in a J1/J3 +π mirrored pose that looks identical to the eye. Align to the robot's home before enabling control, every time.
A torque warning that reads like a power fault
For passive reading, USB power is enough. I first took a torque-enable warning to mean the bus was under-powered and bought an external 5V supply; it only matters if you actually drive the servos.
Wrong Dynamixel cable type
XL330s use a specific JST connector/pitch. A lead from another Dynamixel family may fit but mis-seat — keep the cable matched to the servo.
Printed holes too tight
3D-printed seats and bores come out undersized. Ream to spec rather than forcing parts — over-pressing cracks mounts and binds joints.
Spring too stiff / wrong angle
An over-tensioned or badly-anchored gravity spring makes the arm fight you or droop. Tune pretension and mount angle so it floats neutrally.
The two classics — eight servos sharing factory ID 1, and stock cables too short for the long links — are covered in §4.
06 · First Power-On Checklist
/dev/serial/by-id (and as /dev/ttyACM0). A udev rule that pins it to /dev/gello saves you from floating port numbers later.gello_get_offset.py from a known start pose, then validated against the homed robot, not free-space matching.gello_software derives each joint's offset from a known start pose: hold the leader in that pose and run the script. Offsets come out as multiples of π/2, the calibration resolution. We calibrate from the pose [0, 0, 0, -1.571, 0, 1.571, 0]and then home the robot to it, so leader and follower can be compared joint by joint.
# gello_software: offsets from a known start pose (leader held in that pose) python scripts/gello_get_offset.py \ --start-joints 0 0 0 -1.571 0 1.571 0 \ --joint-signs 1 1 1 -1 1 -1 1 \ --port /dev/gello
The resulting config, kept in our repo and injected into GELLO's port map at runtime (see §5). The offsets, signs, and gripper range are build-specific, so yours will differ:
# registered into gello.agents.gello_agent.PORT_CONFIG_MAP at startup
PORT_CONFIG_MAP["/dev/gello"] = DynamixelRobotConfig(
joint_ids = (1, 2, 3, 4, 5, 6, 7),
joint_offsets = tuple(m * np.pi / 2 for m in (2, 0, 0, 1, 2, 3, 2)),
joint_signs = (1, 1, 1, -1, 1, -1, 1),
gripper_config = (8, 264, 222), # (id, open deg, closed deg)
)07 · Notes for FR3 Teleoperation
The leader's motors are torque-off; it senses your motion and never pushes back. You feel the springs and the arm's own inertia — nothing from the robot or its contacts — so move deliberately, especially near objects. The leader is read synchronously inside the control loop, so a frame or two of lag at 15 Hz is normal, not a fault.
RLinf ships a Cartesian path: forward kinematics on GELLO's joints gives an end-effector pose, the delta against the robot's current tool pose becomes an EE-delta action, and inverse kinematics redistributes it across the real joints. We started there, then added a joint-space path for data collection: a P-controller on the joint error emits an 8-D joint-velocity action [dq(7), gripper] at 15 Hz.
Cartesian path (RLinf GelloIntervention)
GELLO joints ─(offset, sign)─▶ FrankaFK ─▶ EE pose ─▶ delta vs current TCP ─▶ IK ─▶ joint commands
Joint-space path (GelloJointIntervention — what we collect with)
GELLO joints ─(offset, sign)─▶ v = clip(kp · (q_gello − q_robot), ±vmax) ─▶ [dq(7), gripper] @ 15 Hz
kp = 4, one P-gain for all seven jointsWhy the switch: joint velocity is π₀.₅-DROID's native action, so the same 8-D vector that drives the robot is what lands in the dataset, with no IK in between. It also closes a trap of the Cartesian path: because IK redistributes motion, a wrong joint sign can hide behind perfectly good end-effector tracking. Validate signs with whole-arm tracking, and keep the Cartesian path for policies that act in EE-delta space.
Before each session, pose GELLO to match the robot's home, then enable teleop. Enabling with a large mismatch commands a big instantaneous jump. A live readout of leader-vs-robot joint error is worth the ten lines of code: it catches the mirrored poses the eye misses.
08 · Results & Lessons
The build went from printed parts in mid-May 2026 to a calibrated leader on June 16 and the first teleop rollouts (the clips at the top) the next day. The joint-velocity collection pipeline went live on June 22 with pilot datasets (stack a cube, assemble a lamp); on July 3 the offsets were re-derived against the homed robot after a container rebuild dropped the first calibration. Since then the arm has been the lab's demonstration device for the FR3.
250
episodes across 10 tabletop tasks in the lab's FR3 dataset, 66,463 frames
15 Hz
8-D joint-velocity actions, recorded directly as the dataset action
2 × 224²
camera views per frame (LeRobot v2.1), plus full-HD SVO recordings for re-cropping
4–8 mrad
leader-to-robot joint error on J1–J4 at 1–2 frames of lag (measured Nov 2026)
Tasks range from pick-and-place and stacking to pressing a button, closing a shape-sorter lid, inserting a book into a stand, and pulling a block out from under another. The dataset was collected by lab members with this leader arm and the joint-velocity pipeline in §7.
franka_fr3 set used here.PORT_CONFIG_MAP, and gello_get_offset.py.How the rest of the cell is wired — the FR3 controller on the NUC, the cameras, and the zero-shot π₀.₅-DROID evaluations — is documented on the FR3 cell page →