Pelican-VLA 0.5: Attending Before Acting Benefits Generalization

Zeyuan Ding; Wenhai Liu; Yang Xu; Jiayu Hu; Yinda Chen; Yi Zhang; Yong Dai; Jian Tang; Xiaozhu Ju

Pelican-VLA 0.5: Attending Before Acting Benefits Generalization

Robotics 2026-07-07 v1 Machine Learning

Authors: Zeyuan Ding , Wenhai Liu , Yang Xu , Jiayu Hu , Yinda Chen , Yi Zhang , Yong Dai , Jian Tang , Xiaozhu Ju

Abstract

In this report, we present Pelican-VLA 0.5, a unified VLA model that integrates vision-language understanding, future-frame generation, and action prediction within a single architecture. Pelican-VLA 0.5 achieves attention-level generalization: without object annotations, segmentation masks, attention supervision, or task-specific fine-tuning, its action pathway already focuses on the instruction-relevant object and contact region. This behavior persists across unseen scenes and unseen robot embodiments, and is substantially stronger than in other open-source VLA baselines. We verify that this ability originates from the learnable Reasoning Slots inserted between perception and action: by routing task-relevant visual information through a compact bottleneck, the slot interface induces manipulation-centric attention during pre-training and remains effective across different policy structures, including a MoT-style architecture.

Keywords

vision-language-action model generalization attention reasoning slots robot manipulation pre-training

Cite

@article{arxiv.2607.06655,
  title  = {Pelican-VLA 0.5: Attending Before Acting Benefits Generalization},
  author = {Zeyuan Ding and Wenhai Liu and Yang Xu and Jiayu Hu and Yinda Chen and Yi Zhang and Yong Dai and Jian Tang and Xiaozhu Ju},
  journal= {arXiv preprint arXiv:2607.06655},
  year   = {2026}
}