Pelican-VLA 0.5 teaches robots to 'attend before acting,' boosting generalization without extra training

Pelican-VLA 0.5, introduced by Zeyuan Ding and colleagues, is a unified Vision-Language-Action (VLA) model that integrates vision-language understanding, future-frame generation, and action prediction within a single architecture. Its core innovation is the insertion of learnable 'Reasoning Slots' between perception and action modules. These slots act as a compact bottleneck that compels the model to route only task-relevant visual information, which induces manipulation-centric attention during pre-training. Remarkably, this attention-level generalization emerges without needing object annotations, segmentation masks, attention supervision, or task-specific fine-tuning. The model's action pathway automatically focuses on the instruction-relevant object and contact region, and this behavior persists across unseen scenes and unseen robot embodiments. The architecture also works with different policy structures, including a MoT-style (Mixture of Transformers) design. Compared to other open-source VLA baselines, Pelican-VLA 0.5 shows substantially stronger attention generalization, making it a promising step toward more sample-efficient and adaptable robot learning.

The practical implications are significant. By achieving attention-level generalization out of the box, Pelican-VLA 0.5 reduces the need for costly manual annotations and task-specific fine-tuning for each new robot or environment. This could accelerate deployment of robotic systems in unstructured settings like homes or warehouses, where object labels and segmentation masks are rarely available. The paper (arXiv:2607.06655) contributes to both robotics and machine learning, suggesting that a well-designed bottleneck between perception and action can enable models to 'attend before acting' without explicit supervision. Future work may extend these findings to more complex manipulation tasks and real-world physical robots. For now, Pelican-VLA 0.5 offers a tangible path toward more generalizable and autonomous robotic agents that can interpret language instructions and act correctly in novel situations.