A new paper on arXiv presents a geometry-preserving approach to retargeting upper-body human motion from monocular RGB video to robots. The work aims to make single-camera footage a viable source of demonstrations for robot programming.
The authors note that video-driven transfer is challenging because body and hand motion are recovered with different spatial characteristics. This mismatch complicates direct mapping from human poses to robot configurations. The proposed method focuses on preserving geometric consistency throughout the retargeting process.
By addressing the spatial discrepancy between body and hand estimates, the approach could improve the reliability of learning from video. The paper is available on arXiv with the identifier 2609.37776.