GELLO Teleoperation System Overview

• GELLO is a low-cost, open-source teleoperation system that uses a kinematically isomorphic, 3D-printed controller arm to mirror robot joint movements.
• It maps operator motion directly to robot joint coordinates via calibrated joint-space transformation, eliminating the need for complex inverse kinematics.
• User studies and system extensions demonstrate GELLO’s high success rates in contact-rich and bimanual tasks, making it pivotal for imitation learning research.

Searching arXiv for the specified GELLO paper and related follow-up work to ground the article in current literature. GELLO, short for GEneraL, LOw-cost, and intuitive teleoperation, is a teleoperation framework for robot manipulators built around a small controller arm that is kinematically equivalent to the target robot. Rather than commanding the follower in Cartesian end-effector space, GELLO maps the operator’s motion directly into the robot’s joint coordinates through an isomorphic arm made from 3D-printed parts and low-cost DYNAMIXEL servos, with a bill of materials under \$300 per controller arm. The original system was introduced as an open-source platform for the UR5, xArm7, and Franka Panda, with the explicit goal of lowering the barrier to collecting large, high-quality demonstrations for imitation learning (Wu et al., 2023). Subsequent work has treated GELLO both as a baseline and as an extensible substrate, including a Global–Local interface for coarse-to-fine teleoperation near singular configurations (Zhou et al., 14 Feb 2025) and a force-augmented variant that reflects follower-side interaction torques and incorporates force signals into policy learning (Sujit et al., 18 Jul 2025).

1. Motivation and problem setting

The central problem addressed by GELLO is the data-collection bottleneck in imitation learning. The original paper frames teleoperated demonstrations as a prerequisite for acquiring complex manipulation skills and notes that policy quality is constrained by the quality, scale, and variety of the demonstration corpus. It further situates this issue in the broader trend that performance and generalization continue to improve as dataset scale grows, citing prior work such as RT-1 and RT-2. In this setting, teleoperation is not merely a control interface; it is an upstream determinant of what kinds of behaviors can be captured at all (Wu et al., 2023).

The system is motivated in part by the limitations of common low-cost teleoperation devices. VR controllers and 3D spacemice operate in abstracted end-effector space and do not expose the operator to the robot’s joint-space constraints. As reported in the GELLO paper, this means that users cannot directly sense or anticipate singularities, joint limits, or self-collisions. VR- and camera-based interfaces may additionally suffer from occlusions and added latency. At the other end of the design spectrum, haptic devices and robot-to-robot bilateral setups can mitigate some of these issues, but are described as orders of magnitude more expensive and often bespoke to a particular robot arm.

GELLO’s response is to preserve kinematic structure at the human interface. The operator manipulates a scaled arm with the same joint topology and axis arrangement as the follower, so the user’s proprioception carries information about reachable configurations and unfavorable postures. This suggests that GELLO is best understood not simply as a cheap master device, but as a joint-space teleoperation paradigm in which the interface itself encodes the manipulator’s kinematic constraints.

2. Mechanical architecture and low-cost design

The hardware architecture is organized around kinematic isomorphism: for each supported target arm, the GELLO controller has the same number of joints, the same joint topology, and a one-to-one correspondence between controller links and robot links. The original system was built for three commonly used platforms (Wu et al., 2023).

Target robot	DOF / topology	GELLO support
UR5	6-DOF, 6R	Designed and open-sourced
xArm7	7-DOF, 7R	Designed and open-sourced
Franka Panda	7-DOF, 7R	Designed and open-sourced

The controller arms use ROBOTIS DYNAMIXEL XL330 servos, described as low-cost and backdrivable, with 12-bit position encoders and a built-in communication protocol. In the reported design, these servos primarily provide accurate joint-angle readout and passive resistance. Their passive resistance when backdriven contributes natural damping to user motion. The load-bearing structure is made from custom 3D-printed links and brackets, allowing inexpensive and rapid adaptation to different arm geometries.

A further design element is passive gravity compensation. Rubber bands or springs are added to the joints most affected by gravity in the resting pose. For a UR-style layout, the second and third joints receive this compensation. The stated purpose is to stabilize the controller’s natural posture and reduce the tendency to fall into unfavorable elbow-down configurations. The original system also supports bimanual teleoperation by operating two GELLO controllers simultaneously to drive two robot arms.

The system’s cost structure is central to its identity. The bill of materials is reported as less than \$300 USD per GELLO controller arm**. The paper gives approximate comparisons of **~\$150 for a SpaceNavigator 3D mouse, ~\$300** for a Meta Quest 2 VR setup, **~\$30,000 for robot-to-robot teleoperation using an extra UR5, and ~\$40,000 for a Force Dimension Omega7. Within that comparison, GELLO occupies a position close to commodity consumer interfaces in price while retaining robot-specific kinematic structure.

3. Joint-space mapping, calibration, and operation

GELLO’s teleoperation path is defined at the joint level. If the target robot configuration is denoted by $q_r$ and the controller configuration by $q_c$, the reported mapping is

$q_r = S q_c + b,$

where $S \in \mathbb{R}^{n \times n}$ is typically diagonal with entries $\pm 1$ to account for axis direction, and $b \in \mathbb{R}^n$ accounts for zero-offset differences between controller and robot home poses (Wu et al., 2023). Because the controller and target arm share the same kinematic structure, the main teleoperation path does not require inverse kinematics.

The paper also provides a generic least-squares calibration formulation:

$$\min_{S,b} \sum_i \left\lVert q_r^{(i)} - S q_c^{(i)} - b \right\rVert_2^2.$$

If finer structural calibration is desired, an optional Cartesian alignment objective can be posed in $SE(3)$:

$300** for a Meta Quest 2 VR setup, **~\$0

with

$300** for a Meta Quest 2 VR setup, **~\$1

The paper is explicit, however, that this is a generic option rather than the core GELLO teleoperation mechanism. Conceptually, the robot forward kinematics is $300** for a Meta Quest 2 VR setup, **~\$2 and the controller forward kinematics is $300** for a Meta Quest 2 VR setup, **~\$3, but the teleoperation signal is transmitted as direct joint commands.

On the follower side, if the robot exposes joint-space position control, the paper gives a standard PD tracking law of the form

$300** for a Meta Quest 2 VR setup, **~\$4

An implementation-compatible operation sequence is also described: initialize communication with the DYNAMIXEL servos and the robot API; align controller and robot home configurations; estimate $300** for a Meta Quest 2 VR setup, **~\$5 and $300** for a Meta Quest 2 VR setup, **~\$6; continuously read controller joint positions; compute $300** for a Meta Quest 2 VR setup, **~\$7; validate limits; publish the joint command; and stop on collision or fault. Operators are instructed to avoid self-collisions and environment collisions, and the robot is stopped immediately on collision or failure during experiments. Specific middleware, sampling frequencies, latency figures, controller gains, and exact gripper-control details are not reported in the original manuscript.

4. User study and measured performance

The main empirical evaluation in the original GELLO paper is a 12-participant user study with brief orientation and practice. Participants received 6 minutes of orientation covering the robot, teleoperation basics, and task requirements, followed by 5 minutes of practice per device. The study used a bimanual station with two UR5s and five tasks: placing a hat on a rack, opening a case and fetching a sleeping mask, handing over a banana to a kitchen area, folding a towel, and plugging in a USB cable. Task limits were 45 seconds for the hat task and 90 seconds for the remaining tasks. The devices compared were GELLO, VR controllers using Meta Quest 2, and 3D mice (SpaceMouse), with VR and 3D mouse sensitivity tuned to trade off fine USB insertion against faster traversal (Wu et al., 2023).

The reported average success rates were 0.92 for GELLO, 0.72 for VR, and 0.63 for 3D mice. GELLO’s taskwise success rates were Hat 0.92, Mask 0.92, Banana 1.00, Towel 0.92, USB 0.83. For 3D mice, the corresponding values were 0.75, 0.58, 0.67, 0.58, 0.58; for VR, they were 0.92, 0.83, 0.75, 0.58, 0.50. The paper states that GELLO consistently achieved the highest success rates across tasks, with particularly large margins on bimanual and contact-rich tasks.

For successful trials, GELLO is reported to have shown consistently faster completion times than VR and 3D mice across all tasks. Exact timing statistics, including means, standard deviations, and $300** for a Meta Quest 2 VR setup, **~\$8-values, are not provided. Failure counts are aggregated over timeout, self-collision, environment collision, and other irrecoverable failures; within that aggregation, GELLO is reported to have the lowest timeout count and lowest collision risk, although detailed numerical counts are not reported.

Beyond the controlled study, the paper demonstrates GELLO on contact-rich tasks, including filling water bottles requiring robot payload capability, as well as long-horizon tasks, cluttered scenes, and bimanual coordination such as towel folding and banana handover. It also reports operation with 7-DOF arms including the Franka Panda and xArm7, where direct joint mapping allows users to manage null-space motion intuitively during cluttered manipulation. A plausible implication is that GELLO’s value increases with kinematic redundancy, because the operator remains in direct control of null-space choices rather than delegating them to an inverse-kinematics solver.

5. Variants and later extensions

A substantial later extension is the Global–Local (G–L) Teleoperation Interface, which explicitly augments a GELLO-style arm-replica master with a second, local fine-manipulation channel. In that framework, the global component remains a scaled arm replica for efficient workspace traversal and collision-aware positioning, while the local component is implemented using unconstrained devices such as a 6-DoF haptic stylus or IMUs, enabling scaled micro-motions and fine dexterity near the end effector (Zhou et al., 14 Feb 2025).

Two decouplings are introduced. In temporal decoupling, global and local components control the arm sequentially via foot pedals; while local mode is active, the global replica mirrors the slave to preserve continuity. In spatial decoupling, the first $300** for a Meta Quest 2 VR setup, **~\$9 joints are controlled by the global replica and the last three joints by a local IMU-based orientation interface, allowing simultaneous position and orientation control. The paper further distinguishes Direct Mapping (DM), used when the Jacobian is well conditioned, from Singularity-Avoiding Mapping (SAM), which uses damped least-squares or an optimization-based QP near singularities or poor conditioning. It defines singularity metrics such as the minimum singular value $q_r$0, manipulability $q_r$1, and condition number $q_r$2, and also introduces an operational Jacobian for the local loop, defined under DM as $q_r$3 and under DLS-based SAM as $q_r$4.

In that study, the global-only mode of the G–L system is treated as equivalent to a GELLO baseline. Quantitatively, GELLO scored 3/7 on Nested Bucket Picking, 7/7 on Pencil Picking, 5/7 on Membrane Touching, 6/7 on Buckle Opening, and 7/7 on Handle Lifting, while the G–L system achieved 6/7, 7/7, 7/7, 7/7, 7/7 on the same tasks. The paper states that the G–L interface outperformed GELLO on most completion-time measurements except Nested Bucket Picking, but exact times are not reported. These results characterize an augmentation of GELLO rather than the 2023 base system.

Another extension adds force information to GELLO. In the force-augmented variant, implemented with a Franka Emika Panda as the follower, the leader still transmits joint positions, but the follower’s estimated external joint torques are reflected back to the leader as motor currents. The conceptual relation between end-effector wrench and joint torque is written as

$q_r$5

and the leader torque reference is

$q_r$6

The implementation uses the Franka Control Interface for $q_r$7, DYNAMIXEL current control on the leader, and a 50 Hz control and logging loop (Sujit et al., 18 Jul 2025). The work emphasizes that this is a pragmatic force–position scheme rather than full four-channel bilateral control, because the low-cost leader cannot support the dynamic identification and sensing requirements of matched bilateral teleoperation.

The same paper also studies force as an observation channel for imitation learning using Action Chunking Transformer (ACT). The reported setup uses 25 demonstrations per task, with 22 for training and 3 for validation, 3 random seeds per task, 5000 epochs, and 15 rollouts per trained policy, yielding 45 evaluation rollouts per task and per input condition. The ACT chunk length is $q_r$8, and the CVAE regularization coefficient is $q_r$9. Comparing Position only against Position + Force, the reported success rates are: Nut Assembly in simulation, 0.60 ± 0.13 versus 0.42 ± 0.10; Door Opening in simulation, 0.96 ± 0.21 versus 1.00 ± 0.00; Drawer Opening in the real world, 0.62 ± 0.48 versus 0.93 ± 0.25; and Whiteboard Erasing in the real world, 0.24 ± 0.43 versus 0.36 ± 0.48. The paper therefore reports improvement on 3 of 4 tasks when force is included. A separate 20-participant user study found no significant NASA-TLX differences between force and no-feedback conditions, but qualitatively reports that users with robotics experience preferred the force-feedback controller.

6. Limitations, reproducibility, and role in robot learning

The original GELLO system is explicitly unilateral. It does not provide force feedback, and its low-torque servos are used primarily as encoders with passive resistance rather than as high-fidelity haptic actuators. This limits performance on deeply contact-rich tasks where force reflection would be beneficial. Other limitations acknowledged in the original paper include the passive, design-specific nature of gravity compensation; the possibility of unfavorable resting postures or operator fatigue if compensation is insufficient; and the limited scope of the user study, which involved inexperienced users with brief training (Wu et al., 2023).

The published documentation also leaves several implementation details unspecified. The original paper does not report detailed Denavit–Hartenberg parameters, exact joint ranges, controller gains, middleware, sampling frequencies, latency figures, numerical failure counts, or dataset logging specifications. Gripper control is demonstrated through tasks involving grasping and manipulation, but the specific gripper interface is not detailed. These omissions do not prevent replication of the central design, but they do delimit what can be inferred from the paper alone.

Later variants introduce additional constraints of their own. In the Global–Local extension, orientation control is reported to be harder than position control; temporal local scaling can become unintuitive for large rotations; and the IMU-based spatial interface can be affected by skin slip, leading to counterintuitive orientation behavior. In the force-augmented extension, stability is limited by the feedback gain $q_c$0; no passivity observer or controller is implemented; low-cost motors require strong down-scaling and a higher-capacity power supply; and the follower-side torque estimate can produce small nonzero currents in free space due to estimation noise (Sujit et al., 18 Jul 2025).

Reproducibility is nevertheless a defining feature of the GELLO project. The original work states that all software and hardware are open-sourced at https://wuphilipp.github.io/gello/, and later summaries additionally point to community mechanical models. The reported resources include CAD files, controller designs, and software modules for UR5, xArm7, and Franka Panda. The adaptation path for a new robot is also stated explicitly: design 3D-printed links and joints to match the target topology, integrate XL330 servos, add gravity compensation to the most affected joints, calibrate joint offsets and signs $q_c$1, validate limits, connect the software to the target robot API, and configure bimanual operation if needed.

Within robot learning, GELLO’s broader significance lies in shifting teleoperation from a specialized hardware problem toward a reproducible data-interface problem. By preserving joint-space structure at low cost, it reduces the barrier to collecting demonstrations for bimanual, contact-rich, and redundancy-sensitive tasks. A plausible implication is that GELLO’s most enduring contribution is not a single controller design, but a replicable template for coupling human proprioception to robot kinematics in a form that is directly useful for imitation learning.