构建空间智能：World Labs 与 Scenix 合并

Building Spatial Intelligence: World Labs and Scenix Merge

李飞飞 Fei-Fei Li · a16z 播客 · 2026-07-28 · 约 42 分钟 · 原视频 ↗

打开互动全文版（中英对照 + 朗读 + 问答）→

本期速览 · Overview

李飞飞与 Yinuo 讨论空间智能的愿景以及用于训练机器人的现实到模拟再到现实的管道。

Fei-Fei Li and Yinuo discuss their vision for spatial intelligence and the real-to-sim-to-real pipeline for training robots.

要点 · TL;DR

• World Labs 与 Scenix 合并，通过大型世界模型构建空间智能。
  World Labs and Scenix merge to build spatial intelligence via large world models.
• 仿真技术实现反事实推理和可扩展的数据生成，对机器人至关重要。
  Simulation enables counterfactual reasoning and scalable data generation for robotics.
• 机器人系统需要可靠性和效率，通过仿真的可控随机化来实现。
  Robotic systems need reliability and efficiency, achieved via simulation's controllable randomization.

核心观点 · Key points

• World Labs 通过大型世界模型构建空间智能，以理解空间并在其中行动。
  World Labs builds spatial intelligence via large world models for understanding and acting in spaces.
• 仿真对机器人至关重要，能够实现反事实推理和状态空间的系统覆盖。
  Simulation is critical for robotics, enabling counterfactual reasoning and systematic coverage of state space.
• 机器人领域缺乏数据；仿真可规模化生成用于训练和评估的数据。
  Robotics lacks data; simulation scales data generation for training and evaluation.
• 机器人需要可靠性和效率；仿真提供可控的随机化和加速。
  Robots need reliability and efficiency; simulation provides controllable randomization and speed-up.
• 机器人应用应从结构化环境推进到半结构化环境，再到非结构化环境。
  Robotic applications should progress from structured to semi-structured to unstructured environments.
• 机器人基础设施应模型无关和具身无关，以服务于各种机器人。
  Robotic infrastructure should be model-agnostic and embodiment-agnostic to serve diverse robots.

反共识 · Contrarian takes

• 机器人世界模型不需要完美保真度；通过随机化捕捉本质结构即可。
  Robotic world models don't need perfect fidelity; capturing essential structure with randomization suffices.
• 3D 世界模型需要时空一致性，不像视频模型可能违反物理规律。
  3D world models require spatial-temporal consistency, unlike video models that may fail physics.
• 机器人实现人类级功率效率还很遥远，尽管特定任务可能性能匹敌。
  Human-level power efficiency in robotics is far away, though narrow tasks may match performance.
• 机器人评估比训练更难；仿真可实现数量级更快的迭代。
  Robotic evaluation is harder than training; simulation enables orders-of-magnitude faster iteration.
• 机器人应使用专用硬件专注于狭窄任务，而非模仿通用人类形态。
  Robots should specialize in narrow tasks with dedicated hardware, not mimic general human form.
• 即使成熟的语言模型也不能盲目信任；机器人必须开箱即用地可靠工作。
  Even mature language models cannot be blindly trusted; robots must work reliably out of the box.

本期章节 · Chapters（共 28）

• 介绍与World Labs Introduction and World Labs
• 空间智能与机器人收购 Spatial Intelligence and Robotics Acquisition
• Yunu背景与Scenix Yunu's Background and Scenix
• 真实-仿真-真实管线 Real-to-Sim-to-Real Pipeline
• 合作故事 Collaboration Story
• Marble与数据稀缺 Marble and Data Scarcity
• 规模与互补性 Scaling and Complimentarity
• 团队成员介绍 Introduction of Team Members
• 加入World Labs动机 Motivation to Join World Labs
• 机器人基础模型 Foundation Model for Robotics
• 视频模型vs.3D仿真 Video Model vs. 3D Simulation Approach
• Synenix务实方法 Robot Aspirations and Tasks
• 精度vs.创意与仿真保真度 Pragmatic Approach of Synenix
• 仿真与反事实推理 Precision vs. Creativity and Simulation Fidelity
• 仿真优势:可靠与高效 Simulation and Counterfactual Reasoning
• 用例:评估与训练 Simulation benefits: reliability and efficiency
• 平台与生态系统 Use cases: evaluation and training
• 无关具体形态的平台 Platform and ecosystem
• 人形机器人预测与受限部署 Embodiment-agnostic platform
• 经济视角与能效比较 Predictions on humanoids and constrained rollouts
• 机器人vs.软件性能 Economic lens and energy efficiency comparison
• 对战略与仿真的影响 Robotics vs Software Performance
• 与Synnex集成 Impact on Strategy and Simulation
• 地域扩展 Integration with Synnex
• 未来产品成功 Geographic Expansion
• 客户参与阶段 Future Product Success
• 何时联系World Labs Engagement Stage for Customers
• When to Contact World Labs When to Contact World Labs

阅读全文双语转录 →