Nvidia 副总裁 Deepu Talla:机器人不能只懂指令,还需预测行动后果
Nvidia 机器人负责人亲自讲了 VLA 模型的局限:机器人会听指令但不会预判后果,得靠 Cosmos 这类世界模型先模拟再动手。做机器人、具身智能方向的开发者值得听听他的融合路线怎么走。
Nvidia 机器人与边缘 AI 副总裁 Deepu Talla 在 The Information 的 AI Deep Dive 节目中提出,仅靠视觉-语言-动作(VLA)模型不足以让机器人可靠行动,因为机器人还需要预测自身动作带来的物理后果。他认为世界基础模型(如 Nvidia 的 Cosmos)可以让机器人在执行动作前先模拟后果,弥补 VLA 模型“理解指令但不理解物理世界”的短板。
Talla 给出的技术路径是把两类模型结合:VLA 负责把语言指令和视觉感知转成动作,世界模型负责在内部推演动作将引发的环境变化。他明确表示“世界基础模型和动作模型最终会融合在一起,它们需要协同工作”。Cosmos 在这一架构中扮演可预测的物理模拟器角色,为机器人提供训练和推理时的前瞻能力。
这番表态明确了 Nvidia 在机器人领域的路线:不是把 VLA 当终点,而是以 Cosmos 为核心构建“感知-预测-行动”闭环。对机器人开发者和企业来说,这意味着训练数据合成、仿真到真实迁移(sim-to-real)可能更多依赖世界模型管线,而非单纯堆 VLA 参数。
Robots need to do more than understand instructions. They need to predict what happens when they act.
Nvidia Vice President of Robotics and Edge AI Deepu Talla explains why vision-language-action models alone aren't enough, and how world foundation models like Cosmos could help robots simulate the consequences of their actions before taking them.
“World foundation models and the action models, ultimately they're all going to be merging together. They need to work together.”
🚀 Watch the full episode of AI Deep Dive: https://thein.fo/4C0UUcZ
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Nvidia, Cosmos, VLA, world models.
Source attribution: curated from The Information (@theinformation). Cross-source corroboration: 8 independent sources have covered this development (including NVIDIA AI, Jim Fan, rohanpaul_ai), confirming industry-wide signal strength.