Whole-body teleoperation of humanoid robots usually depends on online motion retargeting to translate a human operator's movements into stable robot commands. That retargeting step bridges the morphological gap between human and robot, but it also introduces extra processing delay, which can hurt real-time performance and stability.
The new paper from arXiv proposes a different strategy: instead of computing retargeting on the fly, the system uses learned atomic motion primitives. These primitives are pre-learned building blocks of motion that can be selected and sequenced quickly, reducing the latency that comes from continuous retargeting. The authors argue this also makes the system more robust when the human and robot body shapes differ significantly.
The work is positioned as a step beyond conventional retargeting-based teleoperation, though the abstract does not yet provide quantitative results. The main contribution is the conceptual shift from per-frame retargeting to a primitive-based control scheme that prioritizes speed and stability.