AI Practices AI实践 3h ago Updated 1h ago 更新于 1小时前 49

Beyond VLAs: How World Action Models Reshape Robot Manipulation 超越VLA:世界动作模型如何重塑机器人操作

World Action Models (WAMs) replace vision-language model (VLM) backbones with video world models, enabling robot policies to learn physical dynamics rather than just semantic mappings NVIDIA Cosmos 3, an open omni-model built on a Mixture-of-Transformers architecture, serves as a foundational backbone for post-training WAM-based robot policies WAMs demonstrate superior physical generalization, requiring significantly less task-specific data for zero-shot transfer across new tasks, robots, and en World Action Models (WAMs) 基于视频世界模型而非视觉语言模型(VLMs),通过学习物理动力学实现更强的零样本泛化能力,可迁移到新任务、新机器人和新环境 NVIDIA Cosmos 3是开源Omni-model世界基础模型,采用Mixture-of-Transformers架构,训练数据包含约7.67亿图像、3.48亿视频和800万动作样本 WAM相比VLA的核心优势:从多样化数据中学习物理规律、开放世界泛化能力强、适配新机器人只需少量演示数据 Cosmos 3提供三种尺寸(4B Edge、16B Nano、64B Super),支持从工作站到Jetson边缘设备的多级

70
Hot 热度
72
Quality 质量
68
Impact 影响力

Analysis 深度分析

TL;DR

  • World Action Models (WAMs) replace vision-language model (VLM) backbones with video world models, enabling robot policies to learn physical dynamics rather than just semantic mappings
  • NVIDIA Cosmos 3, an open omni-model built on a Mixture-of-Transformers architecture, serves as a foundational backbone for post-training WAM-based robot policies
  • WAMs demonstrate superior physical generalization, requiring significantly less task-specific data for zero-shot transfer across new tasks, robots, and environments compared to VLAs
  • Cosmos 3 is available in three sizes (4B Edge, 16B Nano, 64B Super) and supports deployment from high-throughput workstations to real-time on-device inference on NVIDIA Jetson hardware
  • The key architectural innovation is joint prediction of video and action, where text uses next-token decoding while continuous modalities (image, video, audio, action) are synthesized through iterative denoising

Why It Matters

This represents a fundamental paradigm shift in robot learning—from policies that merely map language instructions to trajectories (VLAs) toward policies that understand and predict physical world dynamics. For AI practitioners, this means robot manipulation systems can generalize far beyond their training distribution with dramatically reduced data collection costs, accelerating the path from simulation to real-world deployment across diverse robotic embodiments.

Technical Details

  • Architecture: Cosmos 3 uses a Mixture-of-Transformers (MoT) design where an autoregressive transformer handles discrete modalities (text via next-token decoding) while guiding a diffusion transformer for continuous modalities (image, video, audio, action via iterative denoising)
  • Training Data: The foundation model was trained on approximately 767 million images, 348 million videos of real-world dynamics, and 8 million action samples spanning robot manipulation, autonomous driving, camera motion, and egocentric motion
  • Policy Variants: Cosmos3-Nano-Policy-DROID (16B parameters) and Cosmos3-Edge-Policy-DROID (4B parameters) are post-trained for the DROID platform (Franka Panda arm with Robotiq gripper), with the Edge variant enabling on-device deployment
  • Core Mechanism: Unlike VLAs that add an action module to a pretrained VLM, WAMs jointly predict future video states and robot actions, leveraging the backbone's inherent physics prior rather than learning dynamics from scratch during post-training
  • Deployment Flexibility: The model supports tiered deployment from high-throughput workstation serving to real-time inference on NVIDIA Jetson hardware, with measurable policy performance improvements across diverse robot embodiments

Industry Insight

  • The shift from VLA to WAM architectures signals that the future of generalist robotics depends on world models with physical priors rather than semantic understanding alone—organizations investing in robot learning should prioritize building or fine-tuning on video world model foundations
  • NVIDIA's decision to open-source Cosmos 3 in multiple sizes creates a competitive moat around the ecosystem while lowering the barrier to entry for WAM-based policy development, likely accelerating industry adoption of world model approaches
  • The demonstrated ability to adapt to new robots with minimal demonstrations suggests that robot manufacturers and researchers can significantly reduce the data collection bottleneck that has historically limited the scaling of generalist manipulation policies

TL;DR

  • World Action Models (WAMs) 基于视频世界模型而非视觉语言模型(VLMs),通过学习物理动力学实现更强的零样本泛化能力,可迁移到新任务、新机器人和新环境
  • NVIDIA Cosmos 3是开源Omni-model世界基础模型,采用Mixture-of-Transformers架构,训练数据包含约7.67亿图像、3.48亿视频和800万动作样本
  • WAM相比VLA的核心优势:从多样化数据中学习物理规律、开放世界泛化能力强、适配新机器人只需少量演示数据
  • Cosmos 3提供三种尺寸(4B Edge、16B Nano、64B Super),支持从工作站到Jetson边缘设备的多级别部署
  • 已发布Cosmos3-Nano-Policy-DROID(16B)和Cosmos3-Edge-Policy-DROID(4B)两个后训练策略模型,针对DROID平台(Franka Panda机械臂+Robotiq夹爪)优化

为什么值得看

这篇文章揭示了机器人操作领域从VLA到WAM的范式转变,为从业者提供了理解下一代机器人策略架构的关键视角。NVIDIA Cosmos 3的开源为研究者和开发者提供了构建零样本机器人策略的实用基础模型,具有明确的工程落地价值。

技术解析

  • WAM架构原理:WAM用视频世界模型替代VLM作为策略骨干网络,联合预测视频和动作。与VLM仅学习描述世界不同,世界模型学习场景如何演化,使后训练无需从零教授动力学,而是 specialization 已具备物理先验的模型。
  • Cosmos 3模型规格:采用Mixture-of-Transformers (MoT)架构,自回归Transformer处理离散token(如文本),扩散Transformer处理连续模态(图像、视频、音频、动作)。文本通过next-token解码生成,其他模态通过迭代去噪合成。提供4B(Edge)、16B(Nano)、64B(Super)三种尺寸。
  • 训练数据规模:约767M图像、348M真实世界动力学视频、8M动作样本,涵盖机器人操作、自动驾驶、相机运动和第一人称运动等多种场景。
  • 后训练策略模型:Cosmos3-Nano-Policy-DROID(16B)和Cosmos3-Edge-Policy-DROID(4B)针对DROID平台后训练,接收语言指令和多摄像头观测,生成机器人动作轨迹。
  • 部署能力:支持高吞吐量工作站推理和NVIDIA Jetson硬件上的实时边缘推理,显著减少任务特定适配数据需求。

行业启示

  • 机器人策略架构演进:从VLA到WAM的转变标志着机器人操作从"语义理解+动作映射"向"物理动力学建模"的范式升级,行业应关注世界模型在机器人策略中的集成路径。
  • 数据效率与泛化优势:WAM利用预训练世界模型的物理先验,大幅降低后训练数据需求,为机器人规模化部署提供可行路径,建议团队优先评估基于世界模型的策略架构。
  • 开源生态机遇:NVIDIA Cosmos 3开源为研究者提供了强大的基础模型,结合其多尺寸选项和边缘部署能力,可降低机器人AI研发的硬件和算力门槛,加速创新迭代。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Robotics 机器人 Multimodal 多模态 Open Source 开源 Research 科学研究 Training 训练