This article is an exclusive interview with Jie Tan, a senior research scientist in Google DeepMind's robotics team. Jie Tan shares his experience transitioning from Computer Graphics to robotics and points out that Reinforcement Learning is the first paradigm shift in the field of robotics, and Large Language Models (LLMs) are the second. He elaborates on how LLMs give robots common sense and language understanding capabilities, giving them a 'brain.' The article delves into whether robot base models should be independent of LLMs. Jie Tan clarifies that they are currently extensions of existing large models, not yet independent disciplines. The biggest challenge is data scarcity, and he proposes a 'Data Pyramid' theory that includes internet data, video data, simulation data, and proprietary robot data, as well as generative AI simulation data as a scalable solution. He also introduces two major breakthroughs in Gemini Robotics 1.5: introducing 'thinking' capabilities into VLA Models to break down complex tasks and achieving cross-embodiment data transfer through 'Motion Transfer,' significantly improving generalization capabilities. Finally, Jie Tan envisions the potential of World Models as a future robot architecture and emphasizes the importance of scalable data.