Introducing Universal Manipulation Exoskeleton (UME)
A low-cost exoskeleton with real-time haptic torque feedback for learning autonomous policies that perform highly force-mediated, tightly space-constrained, visually occluded, whole-body, and long-horizon mobile manipulation tasks.
Using UME, the teleoperator can unsheathe a heavy metal sword completely blindfolded.
https://ume-exo.github.io/
🧵1/N
Positive users praise the new low-cost universal manipulation exoskeleton for its real-time torque feedback and comfort as a breakthrough for force-mediated robotic tasks, while negative users worry about safety hazards and OSHA compliance.
Better robot data is everything. So this looks amazing
Introducing Universal Manipulation Exoskeleton (UME)
A low-cost exoskeleton with real-time haptic torque feedback for learning autonomous policies that perform highly force-mediated, tightly space-constrained, visually occluded, whole-body, and long-horizon mobile manipulation tasks.
Using UME, the teleoperator can unsheathe a heavy metal sword completely blindfolded.
https://ume-exo.github.io/
🧵1/N
The sole purpose of UME’s mechanical design is to maximize comfort and range of motion (RoM) to the person wearing it. The actuators are arranged such that the rotational axes of the motor intersect at the center of the ball joints (shoulder, wrist).
This design gives the highest freedom and consistency to the operator and removes all discomfort like dragging, or parts blocking your movements.
🧵3/N
@litian_liang Training robot policies with only force and torque data. Great for space constrained and visually occluded scenarios where vision is limited
Cost is $1900👌
By capturing the exact forces required for a task, UME data directly unlocks autonomous, compliant behavior in real time.
🧵6/N
One UME, infinite robots.
Our retargeting algorithm bidirectionally maps kinematics and dynamics, allowing UME to seamlessly teleoperate any robot configuration.
🧵5/N
We push autonomous robot manipulation into more realistic scenarios beyond tabletop grasping, that means:
- Grasping a drink from the fridge side holder, in a REAL office kitchen.
- Reaching for a GPU dropped in between 2 half-built desktop PCs.
- Making maximal space on the workbench by shoving the box all the way to the end.
- Flipping a box upright by mediating the contact-force while on a mobile base.
What UME enabled: for each of the above tasks, 1 person collects data with 1 robot, model trains overnight, the autonomous model works tomorrow morning.
🧵2/N
UME provides real-time transparent torque feedback.
This allow the human to demonstrate how the robot can gently squeeze into narrow spaces without damaging the robot or the objects and obstacles.
Then an autonomous policy can be learned from this data.
🧵4/N
@drjingxi We will be open-sourcing the entire project soon to help reproducing this system, including:
- exoskeleton 3D CAD files - exoskeleton bilateral kinematics & dynamics retargeting - robot control software - model training & evaluation
This work is done in Inclusion AI lab at Ant Group, advised by @drjingxi and Professor Mark Cutkosky from Stanford BDML lab.
Website: https://ume-exo.github.io/ Paper: https://arxiv.org/abs/2606.14218
🧵7/N
@chris_j_paxton Looks not OSHA compliant at all though 😭
@litian_liang very cool!
@litian_liang next you must commit robo-seppuku
@litian_liang this is super cool
@litian_liang 🔥🔥
@litian_liang Thought the sword was going to cut that guy in half...
@litian_liang Are you going to publish the STLs and code? I'd love to publish it on http://orobot.io
@chris_j_paxton Banger. Torque feedback is the cheat code
@chris_j_paxton Over time it will be seamless with human movement.
@chris_j_paxton I’m getting tired of seeing kids get kicked by robots and other safety hazards that are totally unnecessary you can do cool stuff and not risk hurting people
@litian_liang Congrats! Very cool work