Imitation learning lets robots acquire skills from human demonstrations, but collecting those demonstrations often requires teleoperation systems that are labor-intensive to use. A new arXiv paper, 'Trajectory-based Diffusion from One-shot Human Video for Generalizable Manipulation,' targets this data-collection bottleneck.

The abstract notes that most existing approaches rely on such teleoperation setups. The paper's title points to a trajectory-based diffusion method that learns from a one-shot human video, with the aim of generalizable manipulation.

The source is a single abstract, so technical details are limited. No comparison with other methods is available in the provided text.