# HOMIE: teleoperating a humanoid with an exoskeleton cockpit instead of motion capture > HOMIE is an open sourced teleoperation system for humanoids, where the operator wears a pair of exoskeleton arms that are isomorphic to the robot, plus a pair of motion-sensing gloves and a pedal for locomotion. The upper-body joint positions are copied directly from the exoskeleton, so no inverse kinematics is needed, and the whole hardware system costs about half a thousand dollars instead of the price of a motion capture rig. **InternRobotics/OpenHomie** — Open-sourced code for "HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit". - Repository: https://github.com/InternRobotics/OpenHomie - Stars: 623 · Forks: 49 - Language: C++ - License: NOASSERTION - Published: 2026-10-22 · Updated: 2026-10-22 · Language: en - Canonical page: https://hysenlabs.com/projects/internrobotics-openhomie ## Isomorphic arms mean the joint angles are already right HOMIE stands for Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit, and the repository is the official implementation of that paper, by Qingwei Ben, Feiyu Jia, Jia Zeng, Junting Dong, Dahua Lin and Jiangmiao Pang, from the Shanghai Artificial Intelligence Laboratory and the Chinese University of Hong Kong. The central idea is the isomorphism. The operator wears a pair of exoskeleton arms that are isomorphic to the robot, meaning the arm's joint angles are the same as the robot's joint angles. The operator's arm is a physical copy of the robot's arm, so when the operator raises their forearm the exoskeleton's forearm angle is the value the robot needs, and no conversion is required. The README is explicit about the consequence. Because each glove has 15 degrees of freedom, which is more than most existing dexterous hands, the upper-body joint positions can be set directly from the exoskeleton readings, dispensing with IK and achieving faster and more accurate teleoperation. That is the whole design argument, and it is a good one: the usual pipeline is capture a human motion, retarget it onto a different robot skeleton, solve inverse kinematics for the hand, and the errors compound. The gloves can be detached from the arms, which allows them to be reused in systems isomorphic to different robots. That is a small design choice with a large practical effect, since the gloves are the expensive, fiddly part and the arms are the part that has to match the robot. The cost is stated as a number. The total cost of the hardware system is only 0.5k dollars, significantly lower than MoCap devices, and the README connects that to the paper's claim that unlike previous whole-body control methods that depend on motion priors derived from motion capture data, this framework eliminates the dependency, resulting in a more efficient pipeline. ## A pedal for locomotion leaves the upper body free The hardware system is described as three things: the isomorphic exoskeleton arms, a pair of motion-sensing gloves, and a pedal. The pedal is the part that makes the cockpit work as a whole rather than just for the arms. The README says the pedal design for locomotion command acquisition liberates the operator's upper body, enabling simultaneous acquisition of upper-body poses. Without it, walking would need the operator's hands or a separate controller, and the upper-body teleoperation that the exoskeleton provides would be unavailable at the same time. So the division of labour is: the exoskeleton arms and gloves carry the upper body, the pedal carries locomotion, and the operator can do both at once. That is a simple division and it is the reason the design can use an isomorphic arm rather than a more general capture device. The paper is on arXiv at 2502.13013, with demos on YouTube and a project page. The news section records three events, the paper and demos in February 2025, the release of all HOMIE resources in the same month, and then a July 2026 clarification that there is a HOMIE v2 reinforcement learning policy in the NVlabs GR00T-WholeBodyControl repository under decoupled_wbc, which the author says was made for some demos at Gear, with a recommendation to use the latest SONIC model directly. That clarification is worth reading carefully if you are following the project. It points at a different repository for the current policy, which suggests the training code in HOMIE itself is not where the most recent work lives, and it is the difference between the v1 policy described in the paper and what the project recommends now. ## Three sub-repositories that can be taken independently The repository contains three key components, and the README asks you to read them as three independent repositories, each with its own README describing its usage and functions. HomieRL is the reinforcement learning training framework, and its job is to enable different kinds of humanoid robots to walk and squat robustly under any continuously changing upper-body poses. That phrase is the hard part of the problem, since the upper body is being driven by a human at the same time as the lower body is learning to balance. HomieHardware contains the files needed to reimplement the hardware system: design files, PCB principle files and Keil code for the PCBs. The access condition is strict. You are required to fill in a form to get such resources, and the README asks that you use your actual name and actual institution, saying that people without a valid name or institution will not be considered appropriate to have access, in order to protect the knowledge privilege. Individuals without any institutions may contact the authors directly. HomieDeploy contains the deployment code for both the PC connected to the hardware system and the Unitree G1 with Dex3 hands. That is the target hardware in the prerequisites, so the deployment side is written against a specific robot. The training side contributes three techniques to the reinforcement learning framework, and they are worth naming because they are what the paper claims as novel: an upper-body pose curriculum, a height tracking reward, and symmetry utilization. Together they are what allow the robot to walk, squat quickly to any required height and keep its balance while the upper body is moving, which is what expands the operational workspace. The repository listing backs up the three-part structure and adds a fourth directory that the README does not describe, MujocoDeploy, which suggests a MuJoCo deployment path alongside the IsaacGym one. ## The environment is specific, and the hardware is a Unitree G1 The prerequisites list is short enough to be copied into your own plan, and it pins the software stack tightly. The operating system recommendation is Ubuntu 20.04 or 22.04. The simulation environment is IsaacGym Preview 4.0, which needs an NVIDIA GPU, an RTX 2070 or higher, and a driver at version 535.183 as recommended. Conda is the environment manager, with Python 3.8 inside it. The hardware is a Unitree G1 with Dex3 hands, plus one Realsense D455 and two Realsense D435 cameras. The camera arrangement is worth noting: a different model for the one that is presumably the main view and a matched pair for something else, and the project does not say which is which in the visible text. The installation step is thin, because the repository is three projects rather than one. You clone the repository to your Ubuntu computer, and then follow the README in each sub-repository to install all three parts or just one of them. That is the practical benefit of the split, and it is why the split was made. The clone URL in the README is a plain text block, which is itself a small sign of how lightly the documentation is edited: ``` https://github.com/OpenRobotLab/Homie.git ``` Note the organisation name in that URL against the repository name in the top level, which is InternRobotics/OpenHomie. The two differ, and it is worth checking which one you actually want, since the clone instructions and the repository you are reading are not the same path. Support is by email to elgceben@gmail.com, and there is a WeChat discussion group that has passed 200 people, so joining now means adding the person by WeChat with a message asking to join. ## The licence forbids commercial use, and the TODO list is honest The licence is the first thing to read for anyone thinking beyond research. The README states that HOMIE is fully open sourced, however it is strictly forbidden to use HOMIE for any commercial purposes, and the licence file itself is a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Licence. The repository licence field is recorded as one GitHub cannot classify, so the LICENSE file and the README statement are the two things to rely on. The attribution and share-alike parts are the usual Creative Commons obligations, and the non-commercial part is the restrictive one. A company that wants to put a humanoid teleoperation cockpit into a product cannot use this code, and the hardware design files are behind a form that asks for a real name and a real institution, which is the authors' way of keeping the mechanical design out of commercial hands. The TODO list is short and marked honestly with checkboxes. Released the paper with demos, and released all necessary code, are both done. Releasing training code for more robots is not done, and neither is upgrading the low-level control policy for more complex terrains and more robots. Those two open items are the honest limits of the release. The framework as published targets a specific robot and a specific set of terrains, and extending the training code to other robots is work the authors have not done. Anyone planning to run this on hardware other than the Unitree G1 should treat that as the first project rather than a configuration change. There is a citation block for the paper, a BibTeX entry pointing at the arXiv preprint, and an acknowledgements section. The license section also points at a Creative Commons URL, which is consistent with the non-commercial terms above. ## Conclusion HOMIE is worth studying if you are building a humanoid teleoperation stack and want to avoid the two usual costs, motion capture hardware and a retargeting pipeline, because copying joint angles directly from an isomorphic exoskeleton removes the inverse kinematics step entirely and the bill for the rig is roughly five hundred dollars. It is the wrong tool if you need a commercial deployment, since the licence explicitly forbids commercial use and the hardware designs are released only to named people at named institutions. Before you build on it, verify three things yourself: that your robot's arm and hand are isomorphic to the exoskeleton in the way the direct joint copying assumes, that the low-level control policy you use matches what the project expects, since the news notes point to a separate whole body control policy and recommend the latest SONIC model, and that the version of the training code you pull supports your robot, because the TODO list leaves training code for more robots as an unfinished item. ## FAQ ### What is HOMIE? It is an official open sourced implementation of HOMIE: Humanoid Loco-Manipulation with Isomorphic Exoskeleton Cockpit, a humanoid teleoperation system from the Shanghai Artificial Intelligence Laboratory and the Chinese University of Hong Kong. It is published as an arXiv paper with demos, and the repository was made fully open source in February 2025. ### How does the exoskeleton cockpit avoid inverse kinematics? The exoskeleton arms are isomorphic to the controlled robot, so the upper-body joint positions are set directly from the exoskeleton readings, which the README says dispenses with IK and gives faster and more accurate teleoperation. Each glove has 15 degrees of freedom, more than most existing dexterous hands, and the gloves can be detached and reused on other isomorphic robots. ### What hardware and software does HOMIE need? The recommended environment is Ubuntu 20.04 or 22.04, IsaacGym Preview 4.0 with an NVIDIA RTX 2070 or higher and driver 535.183, Conda with Python 3.8, and a Unitree G1 with Dex3 hands plus one Realsense D455 and two Realsense D435 cameras. The three sub-repositories can be installed independently. ### Can HOMIE be used commercially? No. The README states that although HOMIE is fully open sourced, it is strictly forbidden to use it for any commercial purposes, and the licence is Creative Commons Attribution-NonCommercial-ShareAlike 4.0. The hardware design files additionally require a form with a real name and institution. ### What are the three techniques HOMIE adds to the reinforcement learning framework? An upper-body pose curriculum, a height tracking reward, and symmetry utilization. Together they let the robot walk, squat quickly to a required height and hold its balance while the upper body is being moved, which is what the paper claims expands the operational workspace beyond earlier whole-body control methods. ## Sources - [InternRobotics/OpenHomie on GitHub](https://github.com/InternRobotics/OpenHomie) - [Issues](https://github.com/InternRobotics/OpenHomie/issues) - [README](https://github.com/InternRobotics/OpenHomie/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/internrobotics-openhomie