Humanoid robots that work in close physical contact with people face a fundamental sensing problem: the partner's body often blocks the robot's view, or the partner moves outside the camera frame. According to a new preprint on arXiv, this severe partial observability breaks standard pose estimation, leaving the robot uncertain about the human's current configuration.
The authors propose a two-part solution. First, a vision-language model guides contact retargeting, using semantic understanding to decide where and how the robot should make contact even when visual cues are missing. Second, a generative partner model produces plausible human poses to fill in the gaps, giving the robot a working estimate of the partner's state during interaction.
The paper is described as a cross-submission announcement, meaning it has been published elsewhere as well. The abstract does not include experimental results or comparisons, so the reported effectiveness of the approach is not detailed in the source material.