Align-Then-StEer (ATE) was proposed by the Embodied Intelligence Team of China Telecom Artificial Intelligence Research Institute (TeleAl) in collaboration with Tsinghua University and other institutions in September 2025. The relevant research results were published in the paper "Align-Then-stEer: Adapting the Vision-Language Action Models through Unified Latent Guidance".

ATE is a novel, data-efficient and plug-and-play adaptation framework that first aligns different action spaces by constructing a unified latent space. Subsequently, a guidance mechanism guides the generation process of diffusion- or flow-based VLAs during fine-tuning, making the output distribution of the model closer to the target domain.