GELLO and Low-Cost Leader Arms: What a Teleoperation Rig Costs

The cheapest part of a data collection rig is the part that decides how good the data is. On an SO-100 setup the leader arm is a second, slightly different copy of the follower, sitting on the desk next to it. The SO-ARM100 bill of materials puts numbers on the split: two arms come to 229.88 USD, a single follower to 121.94 USD. The leader is roughly 108 USD of a 230 USD rig, and it is the only part your hand ever touches.

GELLO, from UC Berkeley, made the same idea explicit for arms costing thousands: a bill of materials under 300 USD that drives a Franka, a UR5 or an xArm7. This article covers what a leader arm is mechanically, what the GELLO study measured, what every cheap leader in the LeRobot ecosystem costs, and what changes inside your recorded episodes when you swap a pad for a leader.

What GELLO is, and what its study measured

GELLO is a general framework for building low-cost, intuitive teleoperation systems, and the idea is one sentence long: build a miniature, kinematically equivalent copy of the target arm from 3D printed links and off-the-shelf servos, read its joint angles, and send them to the real arm as joint commands. The paper is by Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin and Pieter Abbeel, posted 22 September 2023 and revised 18 July 2024.

The reason it exists is the cost table below. Before GELLO, joint-space teleoperation meant a second industrial arm or a haptic device, and the hobby-priced alternatives work in end-effector space, hiding the arm kinematics from you.

The study used two UR5 arms and five tasks: place a hat on a rack, open a case and fetch a sleeping mask, hand a banana across the workspace, fold a towel, plug in a USB cable. Twelve volunteers, none professionally trained, ran all five on all three devices in randomised order.

For imitation learning data the failure breakdown matters more, because each of these is an episode you throw away.

• Timeout, aggregated over all five tasks: GELLO 1, 3D mice 9, VR 5.
• Self collision (follower hitting itself or the other follower): GELLO 1, 3D mice 5, VR 5.
• Environment collision: GELLO 3, 3D mice 6, VR 6.
• Other irrecoverable failures, such as dropping the object out of reach: GELLO 0, 3D mice 2, VR 1.

The three design decisions inside every leader arm

1. The servo is there for its encoder, not its torque

GELLO uses the DYNAMIXEL XL330-M288-T: a 288.4:1 gear ratio, a 12-bit contactless absolute encoder at 4096 pulses per revolution, 0.52 N.m stall torque at 5.0 V, 18 g, and a 3.7 to 6.0 V input range. 4096 counts over 360 degrees is about 0.088 degrees per step, and the paper quotes joint measurements within 0.088 of a mechanical degree.

They picked the highest gear ratio in the family, and not for torque: the motor is never commanded to move. The reduction provides resistance when the operator backdrives it, which the paper calls natural damping that improves stability. A servo is not strictly necessary, the authors note, since only positions are read, but it packages an encoder, a bus protocol and a housing. The Koch v1.1 leader uses the XL330-M077-T at 77.5:1 and 0.215 N.m at 5.0 V: less reduction, lighter feel, less damping.

2. Scale: alpha equals 0.5

GELLO takes the Denavit-Hartenberg parameters of the target arm and scales the link lengths by alpha, noting the optimal value is user and robot dependent and using 0.5. Joint positions are read off the device and sent through as joint commands, with ZMQ between processes. That is the whole control stack: no inverse kinematics, and half size keeps the device portable across the arm's full degrees of freedom.

3. Joint regularization: rubber bands

A passive servo arm cannot hold itself up, so gravity drags the elbow into configurations that are kinematically legal and practically useless, sometimes into the table. GELLO adds rubber bands or springs on the joints that resist gravity hardest, the second and third on the UR design. The paper measured it: without regularization, holding the device stationary takes a roughly constant 1.9 N at every height. With the bands, the restoring force is lower near the table and higher far from it, pushing the operator away from singularities.

ALOHA solved it the same way: an off-the-shelf WidowX with a 3D printed handle-and-scissor mechanism that cuts the force needed to backdrive the motor, plus a rubber band load balancing mechanism against gravity. The stated reason is endurance, making sessions longer than 30 minutes possible.

The family of cheap leader arms, with real prices

Every leader design that publishes a bill of materials, at the cost each project states in its own currency and at its own date. Supplier prices move, so treat these as published figures, not quotes.

The spread between the cheapest leader and the ALOHA leader is a factor of 65, and the cheap end is not obviously worse: U-ARM reports 39 percent higher data collection efficiency than Joycon at comparable success rates. ALOHA is also the system that produced the ACT policy and action chunking, on a rig its authors put at 18,000 USD, or 20,000 with optional add-ons. Cheap leader arms are not a later compromise; ALOHA was itself the cheap option in 2023.

Why the SO-100 leader has its gears removed and the SO-101 does not

This is the same problem GELLO solved with rubber bands. The SO-100 two-arm bill of materials is twelve identical STS3215 servos: the leader is literally a second follower. The follower needs its reduction to hold its own weight; the leader holds nothing and has to be light in the hand. So the official LeRobot assembly page has you remove the gear from all six leader motors before you build the arm, leaving, as it puts it, only the position encoding of the motor and less friction.

The SO-101 replaced that surgery with three factory gear ratios, which is why its README calls the arm easier to assemble with no gear removal. The LeRobot SO-101 page states the follower uses 1/345 gearing throughout, and publishes the leader assignment.

Shoulder lift keeps the full 1/345 reduction because it carries the whole arm against gravity. The wrist joints carry almost nothing and get the lowest reduction, so they feel light where the fine motion happens. That is GELLO's joint regularization argument solved in the parts list instead of with elastics, and the README sells all six as a bundle: three 1/147 (C046), two 1/191 (C044), one 1/345 (C001). For a new build the leader is the strongest argument for the SO-101; the SO-100 against SO-101 comparison covers the rest.

Building and driving a leader arm: the actual commands

These are the current commands from the SO-101 and imitation learning pages of the LeRobot documentation. LeRobot ships hyphenated console scripts (lerobot-teleoperate), and the old python lerobot/scripts/control_robot.py path no longer exists in the repository, so older tutorials will not run against current main.

1. 1
   Install LeRobot with the Feetech SDK

   The extra pulls in feetech-servo-sdk. Without it the SO classes import but cannot open the bus.

   bash

   pip install -e ".[feetech]"

2. 2
   Find the two serial ports

   Unplug the arm when prompted and it reports which port disappeared. Repeat for the second arm.

   bash

   lerobot-find-port
   
   # Linux only, if the port exists but cannot be opened:
   sudo chmod 666 /dev/ttyACM0
   sudo chmod 666 /dev/ttyACM1

3. 3
   Write ids and baud rates into the leader motors

   Brand new motors usually come with a default id of 1, so they cannot share a bus until each has its own. Connect them one at a time, gripper (id 6) back to shoulder pan (id 1). Once per arm, ever.

   bash

   lerobot-setup-motors \
       --teleop.type=so101_leader \
       --teleop.port=/dev/tty.usbmodem575E0031751

4. 4
   Calibrate both arms

   Move the arm to the middle of every joint range and press Enter, setting the homing offsets, then sweep each joint through its whole range. Wrist roll is the exception: a full-turn joint fixed at 0 to 4095 rather than recorded. The id names the calibration file and must match on every later command.

   bash

   lerobot-calibrate \
       --teleop.type=so101_leader \
       --teleop.port=/dev/tty.usbmodem575E0031751 \
       --teleop.id=my_awesome_leader_arm
   
   lerobot-calibrate \
       --robot.type=so101_follower \
       --robot.port=/dev/tty.usbmodem585A0076841 \
       --robot.id=my_awesome_follower_arm

5. 5
   Teleoperate, with the camera streams on screen

   If the follower mirrors the leader and the gripper tracks the trigger, the rig is done. --display_data=true opens a rerun window with camera feeds and joint traces, the cheapest camera placement check you will run.

   bash

   lerobot-teleoperate \
       --robot.type=so101_follower \
       --robot.port=/dev/tty.usbmodem585A0076841 \
       --robot.id=my_awesome_follower_arm \
       --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 30}}" \
       --teleop.type=so101_leader \
       --teleop.port=/dev/tty.usbmodem575E0031751 \
       --teleop.id=my_awesome_leader_arm \
       --display_data=true

6. 6
   Record a dataset

   Same command with a dataset attached. Right arrow or n ends an episode early, left arrow or r re-records it, Escape or q stops and encodes.

   bash

   lerobot-record \
       --robot.type=so101_follower \
       --robot.port=/dev/tty.usbmodem585A0076841 \
       --robot.id=my_awesome_follower_arm \
       --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 1920, height: 1080, fps: 30}}" \
       --teleop.type=so101_leader \
       --teleop.port=/dev/tty.usbmodem575E0031751 \
       --teleop.id=my_awesome_leader_arm \
       --display_data=true \
       --dataset.repo_id=${HF_USER}/record-test \
       --dataset.num_episodes=5 \
       --dataset.single_task="Grab the black cube"

What the leader arm actually changes in your recorded data

This is the part that gets skipped, and the reason a leader arm is worth 108 USD. The ALOHA paper states the rule directly: the joint positions recorded as the action are the leader positions, the human input, not the follower positions. The observation is the follower positions plus the camera feeds. Using leader positions matters, the authors say, because the force applied is implicitly defined by the difference between commanded and actual position, through the controller inside the servo.

Read that twice if you are training a policy that has to press, insert or hold something. Record the follower position and the action equals what happened, so the force signal collapses to nothing. Record the leader position and the action is what the human asked for, with the residual carrying how hard. A pad has no such residual, because no second arm is being backdriven by a wrist. Same task, same success rate, different LeRobot dataset.

Recording defaults that decide your throughput

The other half of data quality is good episodes per hour. The current LeRobot recording defaults:

At those defaults a 50 episode dataset is 100 minutes of wall clock at best, which is why the GELLO timeout counts matter: nine timeouts is nine wasted reset cycles. lerobot-record also refuses to start without a teleoperator, pointing you at lerobot-rollout instead. Recording is defined as a human driving something.

How many you need depends on the model. Here SmolVLA has the lowest floor at 30 episodes, while ACT, GR00T N1.7 and Pi0.5 want at least 50; the rest is on the policies comparison and in how to collect high-quality VLA training data.

Two ways to get from a leader arm to a trained policy

lerobot-record \
    --robot.type=so101_follower \
    --robot.port=/dev/tty.usbmodem585A0076841 \
    --robot.id=my_awesome_follower_arm \
    --teleop.type=so101_leader \
    --teleop.port=/dev/tty.usbmodem575E0031751 \
    --teleop.id=my_awesome_leader_arm \
    --dataset.repo_id=${HF_USER}/so101-cube \
    --dataset.num_episodes=50 \
    --dataset.single_task="Grab the black cube"

Where a leader arm does not help

Four honest limits, because the first three cost people weeks.

• No force feedback, in any design in the table above. GELLO says its motor torque is too low for it, and that bilateral operation would also need force sensing on the target robot. You can watch the gripper deform the object on camera, but your hand does not know.
• Distance. A leader arm works because your wrist is coupled to a device next to the follower; put a network in between and the coupling is gone. Same physics as remote inference here: the control loop is 20 to 485 ms per action step depending on the model, and public-internet round trips on top turn a working policy into a hesitant one. Viable for slow pick and place, not fast reactive motion. See inference latency.
• It does not fix the scene. A perfect leader arm with one badly placed camera produces a policy that fails exactly like a pad-recorded one. The failure mode pages exist because most bad policies are bad data, and most bad data is upstream of the input device.
• GELLO is not for the SO-100. gello_mechanical lists Franka FER, UR5, xArm7, YAM, Lite6, AR4, FR3 and FR3 Duo, and no SO configuration; its bill of materials lives in a Google Doc rather than the repo. If you own an SO-100 your leader is the SO-101 leader, and GELLO is the paper explaining why it is shaped that way.

GELLO is a design pattern first and a set of STLs second; the SO-101 leader, the Koch v1.1 leader and U-ARM are all instances of it. For wider context, the SO-100 complete guide runs assembly through training, and the RoboTurk article covers scaling to many remote humans, the problem a physical leader cannot solve. Here, record your first dataset and the teleoperation docs are the shortest path, and the dataset directory shows what other people's episodes look like.