Universal Manipulation Exoskeleton (UME)
Wearable two-arm exoskeleton that feeds a robot's joint torques back to the operator
License
Apache 2.0
Code and CAD: Apache-2.0, per LICENSE.md in the repository. The bundled OpenArm model keeps its own Apache-2.0 licence file, and the ACT and DETR code carries Facebook's copyright header from the original projects. The Yahboom IMU reader is the vendor's code and states no licence. The project website's template is CC BY-SA 4.0; the paper and videos are the authors'. Checked 2026-11-08 against LICENSE.md and the files in the repository.
Overview
Universal Manipulation Exoskeleton (UME) is an upper-limb exoskeleton, worn on both arms, for teleoperating robot arms and recording demonstrations with real-time torque feedback. Each arm has eight Damiao quasi-direct-drive actuators: seven follow the operator's shoulder, elbow and wrist, and the eighth, at the hand, drives the gripper. The same actuators measure the operator's motion and push back with the joint torques the robot senses, so the operator feels contact, and the recorded data holds whole-arm joint positions and torques. The structure is printed in PLA on aluminium extrusion, an IMU on the back steers a mobile base when the operator leans, and the paper puts the material cost at 1900 USD. UME comes from Ant Group and Stanford University, and its paper is on arXiv. The repository holds the CAD, the MuJoCo model, the teleoperation and recording code and the ACT training scripts used in the paper.
UME is an exoskeleton for teleoperation that closes the loop in both directions. Its joints are low-reduction quasi-direct-drive actuators from Damiao, which can be back-driven by the arm that wears them and can also apply torque to it. While the operator moves, UME commands the follower robot; when the robot meets resistance, its measured joint torques are mapped back onto the exoskeleton joints, with gravity, Coriolis and friction compensation added to make the feedback as transparent as possible. The authors show an operator drawing a sword from its sheath while blindfolded, guided by the torque feedback alone.
The mechanics follow the human arm. The axes of joints J1 to J3 meet at the centre of the shoulder, J4 is the elbow, and the axes of J5 to J7 meet at the centre of the wrist, so the exoskeleton keeps the full range of motion of the wrist and most of the shoulder. Retargeting splits both the exoskeleton and the robot into a virtual spherical shoulder, an elbow and a virtual spherical wrist: forward kinematics gives each sphere's orientation, inverse kinematics sets the robot's matching joints, and the sub-manipulators' Jacobians map velocities forward and torques back. The paper uses this to drive the 7-DoF OpenArm, 6-DoF UFACTORY X-ARM arms, and 7-DoF Franka arms in MuJoCo. A Yahboom IMU on the back gives the frame's orientation for gravity compensation and turns the operator's lean into velocity commands for a mobile base.
The build uses 8 DM-J4340-2EC and 8 DM-J4310-2EC actuators, 4 kg of Bambu Lab PLA Basic, 2020 aluminium extrusion, a dual-channel CAN-FD to USB adapter, an adjustable 30 V 10 A DC power supply, XT30 cables, M3 and M5 screws and a quick-detach belt; Appendix E.2 of the paper prices the whole system at 1900 USD. The hardware guide covers configuring each actuator's CAN IDs and bus speed with Damiao's debugging tool, print settings, assembly tips and joint zeroing, and the software guide covers the CAN set-up, a gravity-compensation check, teleoperation, data recording and training. Printable STL files sit with the MuJoCo model and editable STEP files are in the ume_step folder. The finished exoskeleton weighs about 12 kg, which the authors say a typical adult can wear comfortably for up to 2 hours.
For the experiments the team built a mobile manipulator from a WowRobo OpenArm 1.0 bimanual arm on four Hexfellow PCW-25 powered casters, for 9533.1 USD in parts. They trained ACT policies that take joint torques as an extra input alongside images and joint positions. Over 20 trials per task, UME policies succeeded in 0.90 of box-pushing runs, 0.85 of box flipping, 0.95 of picking a GPU from a gap with about 5 mm of clearance, and 0.95 of a long fridge drink retrieval. The same data without torque scored 0.50, 0, 0.75 and 0.90, and a UMI-style end-effector baseline scored 0.40, 0, 0 and 0. On box flipping, operators collected 3.3 times as many demonstrations per minute with torque feedback as without it. The project website shows the runs on video.
Why it matters
Most low-cost ways to collect robot demonstrations record positions only. Leader-arm rigs such as Mobile ALOHA and passive exoskeletons such as AirExo give the operator no sense of the force the robot meets, and handheld UMI-style grippers record end-effector poses that leave the arm's shape to an IK solver at run time. UME adds the missing channel: the operator feels the robot's joint torques, and the recordings keep whole-arm joint positions and torques, so imitation learning policies can learn when to press harder and how to keep the whole arm clear in tight spaces. In the paper that difference decides tasks such as box flipping, where neither baseline succeeded once. For about 1900 USD in off-the-shelf actuators, PLA and extrusion, with Apache-2.0 CAD and code, it is within reach of labs already building OpenArm-class arms, and its sub-manipulator retargeting is a reusable recipe for cross-embodiment teleoperation between arms with different joint layouts. The evidence comes from one team's mobile manipulator, so independent builds are still to come.
Problem it solves
Teleoperation rigs and handheld grippers used to collect robot demonstrations rarely capture force: the operator cannot feel what the robot touches, so pushes overshoot or stop short, and the policy never learns from torque. UME gives the operator real-time joint-torque feedback through a wearable two-arm exoskeleton and records whole-arm joint positions and torques, so policies can learn compliant, contact-rich behaviour and arm postures that avoid collisions in confined spaces.
Key takeaways
- Wearable two-arm exoskeleton with 16 Damiao quasi-direct-drive actuators that feeds the robot's joint torques back to the operator
- Joint axes meet at the shoulder and wrist centres, keeping the full wrist range of motion and most of the shoulder
- Sub-manipulator retargeting drives 7-DoF OpenArm and 6-DoF X-ARM arms, and Franka arms in MuJoCo
- An IMU on the back drives a mobile base when the operator leans, for whole-body teleoperation
- ACT policies trained with torque reached 0.85 to 0.95 success on four tasks where the no-torque and UMI-style baselines fell short
- About 1900 USD in parts per the paper, with Apache-2.0 CAD, STL and STEP files, print settings and setup guides
Who it is for
- Robot learning labs collecting contact-rich demonstrations for imitation learning
- Owners of OpenArm or other torque-sensing arms with Damiao-style actuators who want force-feedback teleoperation
- Researchers working on bilateral teleoperation, haptics and compliant manipulation
- Teams building bimanual mobile manipulators who need whole-body data collection
Limitations
- The policies are trained on data from one team's in-house mobile manipulator; no outside builds or reproductions are published yet
- Torque feedback needs a follower that measures joint torques; it was shown on real OpenArm and X-ARM arms, while Franka was shown in simulation only
- The exoskeleton weighs about 12 kg and needs a 30 V DC power supply, a CAN-FD adapter and a Linux host, so it is a lab rig rather than a portable kit
- There is no parts table in the repository: costs come from the paper's Appendix E.2, which leaves out the 4040 extrusion, brackets and the actuator configuration adapter shown in the CAD and guide
- Setup relies on the Windows Damiao debugging tool, hand-edited scripts for zeroing, and a mobile base built from Hexfellow casters for the full demos
- No datasets or trained checkpoints are released, and the repository has had no updates since July 2026