The arXiv paper "KungfuAthleteBot" explores using video as a scalable source for teaching humanoid robots highly dynamic motions, such as martial-arts-style movements. According to the abstract, this direction could reduce the need for costly, dedicated data collection, but it introduces several technical hurdles.
The authors identify specific challenges in this video-based learning pipeline: the robot's height can "float" relative to the ground, and the video lacks direct kinematic and force measurements. These missing data make it difficult to transfer observed motion into stable, physically plausible robot control.
The abstract is truncated, so the full details of how KungfuAthleteBot resolves these issues are not included in the summary. Readers interested in the proposed method and its results will need to consult the full paper on arXiv.