智东西2月3日报道，近期，英伟达宣布其全新3D通用模型论文将发表于2026国际3D视觉会议，论文的预印本已于去年7月发表。这篇论文构建出了一种建 构3D世界的新范式，验证了"AI生成的3D合成数据"可规模化替代人工标注数据，能够大幅降低视觉模型预训练的成本。 论文的主要成果为3D-GENERALIST模型，该模型使用统一化框架，将3D环境生成的四大核心要素即布局、材质、光照、资产等统一到序贯决策框架中。 研究团队还提出了基于CLIP评分的自改进微调策略，可以让模型在下一轮生成中能自主修正前序错误。 这篇论文的作者有8位华人，第一二作者都是中国留学生，清华"姚班"出身的斯坦福大学助理教授吴佳俊也名列其中。 CES 2025上，英伟达正式推出世界基础模型平台Cosmos。在CES 2026的演讲中，黄仁勋依旧将"Physical AI"作为了整场发布的核心灵魂，正式将Cosmos 定位为Physical AI的"底层代码"与"世界模拟器"。此外，黄仁勋还发布了Cosmos Reason 2，让AI不仅生成世界，还能用自然语言进行链式因果推理。 3D-GENERALIST这一技术会给英伟达的Cosmos补全哪块拼图 ...

Core Insights - Nvidia is establishing a default platform in the robotics sector, aiming to replicate Android's dominance in smartphone operating systems [1] - The company has released multiple open-source foundational models to enable robots to reason, plan, and adapt across various tasks and environments, all available on the Hugging Face platform [1] - Nvidia's new Jetson T4000 graphics card and the open-source command center OSMO are designed to support the entire robotics development workflow [1][4] - The trend of AI migrating from the cloud to the physical world is evident, driven by decreasing sensor costs, advancements in simulation technology, and improved generalization capabilities of AI models [1][6] Model Matrix Construction - The foundational models released by Nvidia form the core capabilities layer of physical AI [2] Data Generation and Evaluation - Cosmos Transfer 2.5 and Cosmos Predict 2.5 are responsible for data synthesis and robot strategy evaluation, allowing validation of robot behavior in simulated environments [3] - Cosmos Reason 2 is a reasoning-based visual language model that enables AI systems to observe, understand, and act in the physical world [3] - Isaac GR00T N1.6 is a visual language action model specifically developed for humanoid robots, utilizing Cosmos Reason for full-body control [3] - The Isaac Lab-Arena, launched at CES, is an open-source simulation framework hosted on GitHub, addressing industry pain points in robot capability validation [3] Hardware Accessibility - The Jetson T4000 graphics card, part of the Thor series, offers a cost-effective upgrade with 1.2 trillion floating-point AI operations and 64GB of memory, while maintaining power consumption between 40 to 70 watts [4] Strategic Partnerships - Nvidia has deepened its collaboration with Hugging Face, integrating Isaac and GR00T technologies into the LeRobot framework, connecting 2 million robot developers with 13 million AI builders [5] - The open-source humanoid robot Reachy 2 now supports Nvidia's Jetson Thor chips, allowing developers to test various AI models without being locked into proprietary systems [5] - Early signs indicate that Nvidia's strategy is effective, with robotics becoming the fastest-growing category on the Hugging Face platform and Nvidia's models leading in download numbers [5]

Group 1: Core Insights - Nvidia's CEO Jensen Huang announced that the robotics field has entered a "ChatGPT moment" and introduced a series of open-source "physical AI" models [2] - The new AI chips have achieved "full-scale production" with a fivefold increase in computing power compared to the previous generation, specifically designed for AI applications like chatbots [6] - Nvidia's new platform, Vera Rubin, is set to launch in late 2026 and is expected to have a profound impact on the future of AI due to the industry's heavy reliance on Nvidia's technology [10] Group 2: Robotics and AI Models - Huang showcased two robots, BDX and GR00T, demonstrating how they learn and interact with their environment [4] - The Nvidia Cosmos Transfer 2.5 and Cosmos Predict 2.5 models can generate realistic synthetic data for evaluating robot performance in a safe virtual environment [4] - The Nvidia Isaac GR00T N1.6 model allows for precise control of humanoid robots using visual language action capabilities [4] Group 3: AI Chip Advancements - The new AI chip's performance leap is attributed to the proprietary data format developed by Nvidia, which allows for significant performance improvements with only a 60% increase in transistor count [8] - The chip includes a "context memory storage" layer to enhance response times in conversational AI applications [8] - Nvidia has partnered with Groq to strengthen its position in the AI inference market [8] Group 4: Autonomous Driving Initiatives - Nvidia plans to initiate robotaxi trials with partners as early as 2027, showcasing its ambition in the autonomous driving sector [14] - The company has developed a decision-making software named Alpamayo for autonomous vehicles, which records decision processes for engineers to review [12] - Nvidia's Drive AGX Thor onboard computer is priced around $3,500 and is designed to help automakers save development time and accelerate feature deployment [15]