Semantically Structured Mixture-of-Experts for Compositional Robotic Manipulation

Researchers introduced SMoDP, a semantically guided mixture-of-experts diffusion policy for robotic manipulation.

The framework uses VLM-supervised skill predictions and dual contrastive alignment to route action chunks to the right experts by task semantics.

It outperforms prior baselines on benchmarks while improving efficiency, interpretability, and compositional transfer across tasks.