RoboticsRobot learningStory 03
Humans and robots share one teaching glove
What changedSEED-UMI places the same instrumented exoskeleton on a human and a robot, aligning demonstrations and rollouts without open-loop retargeting, segmentation or inpainting. Across five contact-rich tasks, the authors report three times greater demonstration efficiency than exoskeleton teleoperation and 70% average rollout success. Useful results, but still below deployment reliability.
The useful part
Why it matters
Sharing one measurement interface attacks a persistent imitation-learning problem: the human demonstration and robot execution often describe motion differently.
Worth doing
What to do next
Compare calibration effort, operator fatigue and task-level failures against your existing teleoperation data pipeline.
Keep in mind
Good to know
The efficiency and success rates are author-reported on a limited suite, and 70% success is not deployment reliability.
Evidence