AI 技术编年史 2025：空间智能 Spatial Intelligence — 李飞飞的新前沿

发表于 分类于 阅读次数：
2025 年 2 月，李飞飞提出空间智能是 AI 下一 frontier：理解、生成与推理 3D 世界。中英文对照解析架构、趋势与落地场景。

空间智能 Spatial Intelligence：李飞飞提出的 AI 新前沿 | Spatial Intelligence: Fei-Fei Li’s New AI Frontier

English Title: AI Technology Timeline 2025 — Spatial Intelligence

一、背景 | Background

English

In February 2025, Fei-Fei Li articulated Spatial Intelligence as the next major capability beyond language-centric LLMs. While large language models excel at symbols and text, humans and robots operate in 3D space—understanding geometry, depth, affordances, occlusion, and physical interaction. Spatial intelligence means an AI system can perceive, reason about, and act within structured 3D environments, not merely describe them in words.

Li’s World Labs (founded 2024) and Stanford HAI research positioned spatial AI as complementary to world models: language gives abstraction; space gives grounding. The 2025 narrative connected spatial intelligence to AR/VR, robotics, autonomous systems, and scientific visualization—any domain where “where” matters as much as “what.”

Key terms:

Term Definition
Spatial Intelligence Ability to infer 3D structure, relations, and dynamics from 2D or 3D inputs
Affordance What actions an object or surface enables (graspable, walkable, pushable)
NeRF / 3DGS Neural representations reconstructing or rendering 3D scenes from images
Embodied grounding Linking linguistic concepts to physical coordinates and trajectories

中文

2025 年 2 月，李飞飞 将 空间智能（Spatial Intelligence） 定义为超越语言中心 LLM 的下一能力 frontier。大语言模型擅长符号与文本，而人类与机器人在 三维空间 中行动——理解几何、深度、 affordance、遮挡与物理交互。空间智能指 AI 能 感知、推理并在 结构化 3D 环境中 行动，而非仅用文字描述。

Li 创立的 World Labs（2024）与斯坦福 HAI 研究将空间 AI 与 世界模型 互补：语言提供抽象，空间提供接地。2025 叙事将空间智能关联到 AR/VR、机器人、自动驾驶与科学可视化——凡「在哪里」与「是什么」同等重要的领域。

关键词：

术语 定义
空间智能 从 2D/3D 输入推断 3D 结构、关系与动力学
Affordance 物体或表面支持的动作（可抓、可走、可推）
NeRF / 3DGS 从图像重建或渲染 3D 场景的神经表示
具身接地 将语言概念链接到物理坐标与轨迹

Historical arc: Computer vision progressed from 2D classification (ImageNet) → detection/segmentation → NeRF/3DGS reconstruction → 2025 spatial reasoning where models answer relational questions (behind, inside, reachable). Language-only LLMs hit a ceiling on robotics benchmarks until depth-aware encoders became standard in VLMs.

Metric scale 为何 critical： 没有米制尺度，AR overlay 会「飘」；机器人抓取误差 >2 cm 即失败。2025 产品强调 SLAM + learned depth + IMU 融合，误差 <1% 场景深度作为商用门槛。

中文补充： 空间智能与 数字孪生 天然耦合——工厂、城市、手术室先有三维模型，再叠 AI 推理。中国「空间计算」产业政策与 2025 空间智能叙事相互强化，推动 3D 采集设备成本继续下探。

二、架构 | Architecture

English

2025 spatial-intelligence systems typically combine reconstruction, understanding, and generation:

1
2
3
4
5
6
7
8
9
Multi-view images / video / depth / SLAM
              ↓
   3D representation (NeRF, 3D Gaussian Splatting, mesh)
              ↓
   Scene graph + affordance map + semantic voxels
              ↓
   Spatial reasoning module (VLM fine-tuned on 3D QA)
              ↓
   Action: navigation mesh, grasp pose, AR overlay, sim export

Components explained:

  • 3D Gaussian Splatting (3DGS): Real-time radiance fields; 2025 stacks use 3DGS as interchange format between capture devices and AI models.
  • Spatial VLM: Vision-language models augmented with depth tokens, point cloud encoders, or BEV features; answer questions like “Can the robot reach the mug behind the laptop?”
  • Generative spatial models: Text/image → editable 3D scenes (World Labs, Luma, Meta research)—analogous to Sora but output is navigable space.
  • Calibration layer: Camera intrinsics/extrinsics and metric scale so spatial predictions are not merely pretty but measurable.

中文

2025 空间智能系统通常组合 重建、理解与生成：

1
2
3
4
5
6
7
8
9
多视角图像 / 视频 / 深度 / SLAM
              ↓
   3D 表示（NeRF、3D Gaussian Splatting、网格）
              ↓
   场景图 + affordance 图 + 语义体素
              ↓
   空间推理模块（3D QA 微调的 VLM）
              ↓
   动作：导航网格、抓取位姿、AR 叠加、仿真导出

组件说明：

  • 3D Gaussian Splatting： 实时辐射场；2025 栈以 3DGS 作为采集设备与 AI 模型间的交换格式。
  • 空间 VLM： 视觉语言模型叠加深度 token、点云编码器或 BEV 特征；回答「机器人能否够到笔记本后的杯子？」
  • 生成式空间模型： 文本/图像 → 可编辑 3D 场景——类似 Sora 但输出可导航空间。
  • 标定层： 相机内外参与 metric scale，使空间预测 可测量 而非仅美观。

English

Trend Detail
Spatial foundation models Pretrain on indoor scans, street views, sim worlds; unified encoder for robotics + AR
Language ↔ space bidirection LLM plans in words; spatial module grounds to coordinates; feedback updates language state
Single-image to 3D at production quality Feeds e-commerce, game assets, digital twins
Spatial agents Agents that navigate simulators (Genesis) using spatial maps, not just text tools
Privacy-preserving spatial capture On-device 3D reconstruction for home robots without cloud upload

中文

趋势 详情
空间基础模型 室内扫描、街景、仿真世界预训练；机器人 + AR 统一编码器
语言 ↔ 空间双向 LLM 文字规划；空间模块接地坐标；反馈更新语言状态
单图到 3D 量产质量 服务电商、游戏资产、数字孪生
空间 Agent 在仿真器（Genesis）中基于空间地图导航，而非仅文本工具
隐私友好空间采集 端侧 3D 重建，家庭机器人无需上传云端

李飞飞 2025 年公开演讲强调：空间智能不是 3D 版本的 ChatGPT，而是让 AI 具备「在世界里思考」的几何与物理直觉——这与同年 world model、具身智能浪潮同频。

四、优缺点 | Pros/Cons

English

Pros

  • Grounds LLM hallucinations in metric space; reduces impossible robot commands
  • Enables shared world representation across human UI, sim, and real robot
  • Unlocks AR copilots that understand room layout and object permanence
  • Composable with NeRF/3DGS ecosystems already deployed in mapping products

Cons

  • 3D data scarcity vs. text; labeling and capture remain expensive
  • Scale ambiguity from monocular input persists without depth sensors
  • Real-time full-scene understanding still GPU-heavy on edge devices
  • Standard benchmarks (ScanNet, nuScenes) do not capture open-world spatial reasoning

中文

优点

  • 将 LLM 幻觉接地到 metric 空间；减少不可能执行的机器人指令
  • 人机 UI、仿真与真机共享世界表示
  • 解锁理解房间布局与物体恒常性的 AR copilot
  • 可与已用于测绘产品的 NeRF/3DGS 生态组合

缺点

  • 相对文本，3D 数据稀缺；标注与采集仍贵
  • 单目输入尺度歧义仍在，需深度传感器
  • 端侧全场景实时理解仍耗 GPU
  • 标准基准（ScanNet、nuScenes）未覆盖开放世界空间推理

五、应用场景 | Use Cases

English

Scenario Spatial intelligence role
Home robotics Room map, object locations, “put away” tasks with occlusion reasoning
Surgical / medical AR Register instruments to patient anatomy in 3D
Construction digital twin Progress monitoring from site photos → BIM alignment
Autonomous delivery Sidewalk geometry, curb cuts, elevator button localization
Education Interactive 3D explanations of molecules, astronomy, history sites
Game / metaverse Prompt-to-playable level with consistent physics

中文

场景 空间智能作用
家庭机器人 房间地图、物体位置、考虑遮挡的收纳任务
手术 / 医疗 AR 器械与患者解剖 3D 配准
施工数字孪生 工地照片 → BIM 对齐的进度监测
自动配送 人行道几何、缘石坡道、电梯按钮定位
教育 分子、天文、遗址的交互 3D 讲解
游戏 / 元宇宙 提示词 → 物理一致的可玩关卡

六、GitHub 开源生态 | GitHub

English

Repo Role
genesis-embodied-ai/Genesis Physics-accurate sim environments for spatial agent training and validation
openai/sora Spatiotemporal generation informs spatial scene synthesis research
NVIDIA/Cosmos-Tokenizer Tokenizes spatial-temporal data for foundation model pipelines

中文

仓库 作用
genesis-embodied-ai/Genesis 物理精确仿真环境，训练与验证空间 Agent
openai/sora 时空生成研究影响空间场景合成
NVIDIA/Cosmos-Tokenizer 空间时序数据 Token 化，接入基础模型流水线

七、参考资料 | References

  1. Fei-Fei Li — Talks and essays on spatial intelligence (World Labs, 2024–2025)
  2. World Labs — Large world models product direction
  3. Kerbl et al. — 3D Gaussian Splatting for Real-Time Radiance Field Rendering
  4. Chen et al. — SpatialVLM / 3D-LLM benchmark papers
  5. Meta AI — SceneScript and spatial understanding research

八、评估基准 | Benchmarks and Metrics

English

2025 spatial AI teams track:

Benchmark Measures
Spatial VQA Relational QA on ScanNet / 3RScan
Embodied navigation SPL Success weighted by path length in sim
Grasp success @5mm Real robot repeatability with spatial map
3D reconstruction PSNR/SSIM NeRF/3DGS quality vs. laser ground truth
Latency Full-scene refresh rate on AR glasses SoC

中文

2025 空间 AI 团队跟踪：

基准 度量
Spatial VQA ScanNet / 3RScan 关系问答
具身导航 SPL 仿真中路径长度加权成功率
抓取成功 @5mm 真机重复性
重建 PSNR/SSIM NeRF/3DGS vs 激光真值
延迟 AR 眼镜 SoC 全场景刷新率

八、产业观察与深度解读 | Industry Observations and Deep Dive

English

Supply chain and talent: By the second half of 2025, enterprises stopped treating this topic as a pilot KPI and moved it into annual operating plans. Procurement asked for three-year TCO, not demo accuracy. System integrators packaged reference architectures with SLA-backed support, mirroring how cloud migrations matured a decade earlier.

Interoperability: Open APIs (MCP, ONNX, MLIR dialects where relevant) reduced lock-in, but data gravity still tied customers to platforms with the best vertical corpus or compiler backend. Winners combined open runtimes with proprietary gold datasets or silicon-tuned kernels.

Risk register (2025 common items): (1) Evaluation gap—public benchmarks no longer predict production; (2) Security—prompt injection and tool abuse in agentic stacks; (3) Regulatory—algorithm filing, EU AI Act high-risk categories; (4) Talent—shortage of engineers who understand both ML and domain workflows.

Research frontiers carrying into 2026: Tighter world-model / spatial / sim integration; self-evolving alignment with human audit; cross-chip compilers (see 2026 timeline). Teams that invested in measurement—latency, cost per task, failure replay—outperformed teams chasing parameter counts.

中文

供应链与人才： 2025 年下半年，企业不再将此主题仅作试点 KPI，而是写入 年度经营计划。采购要求 三年 TCO，而非 demo 准确率。系统集成商打包 带 SLA 的参考架构，类似十年前的云迁移成熟路径。

互操作： 开放 API（MCP、ONNX、相关 MLIR dialect）降低锁定，但 数据重力 仍把客户绑在拥有最佳垂直语料或编译后端的平台上。胜者 = 开放运行时 + 专有 gold 数据 或 硅片级调优内核。

风险登记（2025 共性）： (1) 评估鸿沟——公开 benchmark 不再预测生产；(2) 安全——Agent 栈提示注入与工具滥用；(3) 监管——算法备案、EU AI Act 高风险类；(4) 人才——既懂 ML 又懂领域 workflow 的工程师短缺。

延续至 2026 的研究前沿： 世界模型 / 空间 / 仿真 更紧耦合；带人工 audit 的 自演化对齐；跨芯片编译器（见 2026 时间线）。投资 度量——延迟、单任务成本、失败回放——的团队胜过追逐参数量。

Glossary reinforcement | 术语 reinforcement

EN 中文 One-line
Foundation model 基础模型 Large pretrained model finetuned for downstream tasks
Finetune 微调 Update weights on domain data
RAG 检索增强生成 Retrieve docs then generate grounded answers
Sim2real 仿真到真实 Transfer policies from simulator to physical world
TCO 总拥有成本 Full cost of ownership over deployment lifetime

总结 | Summary

中文： 2025 年 2 月，空间智能标志 AI 从「会说话」走向「会在世界里想」。它与世界模型、具身机器人形成三角：预测未来、理解空间、执行动作。落地关键在 metric 精度、实时性与跨模态对齐。

English: February 2025 positioned spatial intelligence as AI moving from eloquence to reasoning in the world—forming a triangle with world models and embodied robots: predict, understand space, act. Deployment hinges on metric accuracy, real-time performance, and cross-modal alignment.