Vision-language-action (VLA) models have emerged as the next generation ofmodels in robotics. However, despite leveraging powerful pre-trainedVision-Language Models (VLMs), existing end-to-end VLA systems often lose keycapabilities during fine-tuning as the model adapts to specific robotic tasks.We argue that a generalizable VLA model should retain and expand upon the VLM'score competencies: 1) Open-world embodied reasoning - the VLA should inheritthe knowledge from VLM, i.e., recognize anything that the VLM can recognize,capable of solving math problems, possessing visual-spatial intelligence, 2)Reasoning following - effectively translating the open-world reasoning intoactionable steps for the robot. In this work, we introduce ChatVLA-2, a novelmixture-of-expert VLA model coupled with a specialized three-stage trainingpipeline designed to preserve the VLM's original strengths while enablingactionable reasoning. To validate our approach, we design a math-matching taskwherein a robot interprets math problems written on a whiteboard and pickscorresponding number cards from a table to solve equations. Remarkably, ourmethod exhibits exceptional mathematical reasoning and OCR capabilities,despite these abilities not being explicitly trained within the VLA.Furthermore, we demonstrate that the VLA possesses strong spatial reasoningskills, enabling it to interpret novel directional instructions involvingpreviously unseen objects. Overall, our method showcases reasoning andcomprehension abilities that significantly surpass state-of-the-art imitationlearning methods such as OpenVLA, DexVLA, and pi-zero. This work represents asubstantial advancement toward developing truly generalizable roboticfoundation models endowed with robust reasoning capacities.