Can robots learn shared skills to handle multiple tasks efficiently?

Diffusion policies excel at robotic manipulation but are computationally expensive and struggle to scale across diverse tasks. This work uses a mixture-of-experts approach where a lightweight skill predictor—trained on Vision-Language Model annotations—routes different behavioral phases to specialized experts. By grounding expert assignment in semantic task structure rather than raw latent noise, the method achieves better generalization while using fewer parameters, and transfers effectively to new tasks through efficient fine-tuning.