The paper, posted on arXiv, addresses multitask robotic learning using diffusion policies. It builds on the idea of combining separate factors via linearly composed scores, which allows policies to adapt to new tasks in a modular way.
The authors propose a method for encouraging components to specialize, rather than remain generic. This is intended to improve compositional generalization and task-wise adaptation.
The abstract is truncated, so the full method and experimental results are not available from the source. The description is limited to the approach's motivation and high-level design.