Source-linked AI summary
GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators
Philipp Wu, Yide Shentu, Zhongke Yi, Xingyu Lin, Pieter Abbeel
TL;DR
Imitation-learning systems need larger, more diverse, and higher-quality demonstrations, but collecting them is costly and time-consuming. GELLO addresses this gap with low-cost, kinematically equivalent teleoperation controllers built from 3D-printed parts and economical motors. Its user study reports more reliable and efficient collection than VR controllers and spacemouses, while demonstrations span complex tasks and three robot platforms.
Problem
Imitation-learning performance is bottlenecked by the quality, scale, and variety of human demonstration data, whose collection can be expensive and time-consuming.
Method
GELLO builds low-cost, intuitive controllers with the target arm’s kinematic structure using economical servos, 3D-printed parts, and passive joint regularization.
Results
GELLO consistently achieves the top success rate across distinct user-study tasks and shows effective performance on complex bi-manual, contact-rich, and long-horizon tasks across three robot platforms.
Takeaways & Limitations
GELLO is intended to lower the barrier to collecting large, high-quality demonstration datasets for robot learning.
Takeaways & Limitations
GELLO lacks force feedback because its motors have limited output torque, restricting its capabilities for more contact-rich teleoperation.
Abstract
from arXiv · showhide
Humans can teleoperate robots to accomplish complex manipulation tasks. Imitation learning has emerged as a powerful framework that leverages human teleoperated demonstrations to teach robots new skills. However, the performance of the learned policies is bottlenecked by the quality, scale, and variety of the demonstration data. In this paper, we aim to lower the barrier to collecting large and high-quality human demonstration data by proposing a GEneraL framework for building LOw-cost and intuitive teleoperation systems for robotic manipulation (GELLO). Given a target robot arm, we build a GELLO controller device that has the same kinematic structure as the target arm, leveraging 3D-printed parts and economical off-the-shelf motors. GELLO is easy to build and intuitive to use. Through an extensive user study, we show that GELLO enables more reliable and efficient demonstration collection compared to other cost efficient teleoperation devices commonly used in the imitation learning literature such as virtual reality controllers and 3D spacemouses. We further demonstrate the capabilities of GELLO for performing complex bi-manual and contact-rich manipulation tasks. To make GELLO accessible to everyone, we have designed and built GELLO systems for 3 commonly used robotic arms: Franka, UR5, and xArm. All software and hardware are open-sourced and can be found on our website: https://wuphilipp.github.io/gello/.
I. INTRODUCTION
GELLO addresses the data bottleneck in imitation learning by offering a low-cost, intuitive teleoperation framework designed to make large and diverse demonstration collection more accessible. It implements practical systems for three robot arms and open-sources the hardware and software for replication.
- Imitation-learning performance is bottlenecked by demonstration datasets that lack sufficient size, complexity, and task diversity.
- GELLO builds miniature controllers with kinematically equivalent structures using 3D-printed parts and economical off-the-shelf motors.
- The framework targets scalable demonstration collection by combining affordability, ease of construction, and intuitive human operation.
- GELLO is implemented for three commonly used robot arms and compared with prevalent low-cost teleoperation systems in a comprehensive user study.
- The hardware and software are fully open-sourced with a bill of materials and assembly instructions to support community replication.
A. Teleoperation Systems for Manipulation
Low-cost teleoperation systems reduce expense but often abstract away robot morphology and kinematic constraints. GELLO instead uses a scaled replica of the target arm to retain intuitive, constraint-aware control at lower cost than full robot-arm controllers.
- Commodity devices such as spacemouses, VR controllers, cameras, and IMUs commonly control robots through an abstracted end-effector space.
- Table I compares teleoperation-system costs, with GELLO positioned favorably against spacemouses and VR and far below other options.
- Full-sized isomorphic controller arms preserve robot-like control but can increase cost, while bespoke exoskeletons are typically specific to individual robots.
- GELLO uses low-cost components to create a scaled replica of the target arm while maintaining the advantages of matching kinematics.
- Bilateral teleoperation provides force feedback and can relay the target arm’s environmental forces and kinematic constraints to the user.
B. Learning from Human Demonstrations
Imitation learning benefits from larger datasets, but collecting human demonstrations remains expensive and time-consuming. GELLO complements other human-robot interaction approaches by focusing on accessible and intuitive teleoperation.
- Learning-system performance scales with dataset size, creating sustained efforts to collect larger demonstration datasets.
- Human demonstration collection can be expensive and time-consuming, with one example spanning 17 months and involving a research team.
- Prior approaches include shared control and simultaneous operation of multiple robots, whereas GELLO targets more accessible and intuitive teleoperation systems.
III. TELEOPERATION DEVICE DESIGN
GELLO’s device design combines low-cost servos, a scaled kinematically equivalent structure, 3D-printed parts, and passive joint regularization. These choices support affordable, portable, replicable control across three robot platforms and varied tasks.
- Design principles: GELLO aims to minimize entry barriers through economical backdrivable servos, 3D-printed components, and a minimalist design costing under $300.
- Design principles: The device is designed for complex bi-manual and contact-rich manipulation tasks, while its compact self-contained form supports transport across lab and in-the-wild environments.
- Design principles: Minimal parts sourcing and straightforward assembly make GELLO simple to replicate for multiple robot types.
- Servo selection: DYNAMIXEL XL330 servos provide economical, self-contained joint-position sensing through high-resolution encoders, simplifying construction and maintenance.
- Kinematic structure: A scaled kinematically equivalent structure maps controller joints directly to the target arm, avoiding inverse-kinematics computation and exposing kinematic-limit resistance.
- Joint regularization: Passive springs or rubber bands counteract gravity, preserve a natural posture, reduce collision-prone configurations, and provide force cues near joint-range extremities.
- 3D printed parts: 3D printing enables customization for three kinematically different robots while keeping production costs low.
- Implementations: GELLO was instantiated and tested on the UR5, xArm7, and Franka Panda using direct joint control and robot-specific APIs.
IV. EXPERIMENTS
The experiments evaluate GELLO quantitatively against common low-cost teleoperation systems and qualitatively across multiple robot platforms and manipulation settings.
- GELLO is compared with other common low-cost teleoperation systems across five manipulation tasks.The tasks explore varied manipulation challenges.
- GELLO is also tested on robots from three different manufacturers across diverse manipulation tasks.
A. User Study Procedure
The user study evaluates bimanual teleoperation with 12 inexperienced participants using GELLO, 3D mice, and VR controllers across five varied tasks.
- 12 university volunteers with no professional teleoperation training compared GELLO, 3D mice, and VR controllers.The study used two UR robots under controlled conditions.
- The five tasks included placing a hat, retrieving a sleeping mask, handing over a banana, folding a towel, and inserting a USB cable.The setup used a bimanual station with two UR5 robots.
- The tasks targeted articulated-object interaction, large workspaces, deformable objects, and precise insertion.
- Users received a six-minute general orientation without device-specific instructions before sequentially trying all three devices.
- Task success, failure modes, or completion time were recorded, while direct hand performance provided a human reference.
B. User Study Results
Across the user study, GELLO achieved the highest success rates and consistently faster successful completion times, with fewer expertise- and constraint-related failures than the alternatives.
- GELLO achieved the top success rate across all evaluated tasks.
- In complex banana-handoff tasks, spacemouse and VR use caused more failures, including self-collisions and singularities.
- GELLO had a large success-rate advantage for towel folding and a much higher success rate for USB insertion.For USB insertion, this advantage held despite static-controller strategies available with VR and spacemouse.
- GELLO required minimal user expertise and had the lowest timeout count among the teleoperation systems.
- GELLO’s isomorphic joint structure was associated with the least collision risk, while Cartesian devices commonly produced self-collisions.
- GELLO produced consistently faster completion times than the other devices on successful trials across all five tasks.The comparison excludes unsuccessful trials.
C. Teleoperation System Capabilities
GELLO is demonstrated on challenging real-world manipulation tasks across three robot platforms, including contact-rich, long-horizon, and bimanual coordination tasks.
- GELLO performs contact-rich, long-horizon, and challenging bimanual coordination tasks in real-world environments.
- GELLO supports 7-DOF Panda and xArm robots despite their extra degree of freedom.
- Kinematic equivalence enables users to directly manage the arm’s null space when operating in cluttered spaces.
- Tasks such as filling water bottles require payloads that would be difficult for smaller arms such as the ViperX used in ALOHA.
V. DISCUSSION
GELLO lowers the barrier to collecting large, high-quality demonstration datasets, but its low-cost design omits force feedback and the user study covers only briefly trained, inexperienced users.
- GELLO’s low-cost design omits force feedback, limiting teleoperation for more contact-rich tasks.The authors made this compromise to preserve affordability, accessibility, and applicability across robot arms.
- The user study evaluated inexperienced users after only brief instruction and limited practice.The authors note that additional training can significantly improve proficiency and leave this study for future work.
- GELLO’s demonstrated effectiveness came from a user study of teleoperation with a bi-manual robot system using two UR5s.The platform was also designed for three robots to demonstrate versatility and accessibility.