AI 技术编年史 2025:世界模型 World Model — 机器人与自动驾驶的新范式

世界模型 World Model:机器人与自动驾驶的新范式 | World Models for Robotics and Autonomous Driving

English Title: AI Technology Timeline 2025 — World Models: A New Paradigm for Robotics and AV


一、背景 | Background

English

In early 2025, World Model moved from a niche research term to a central engineering pillar for robotics, autonomous driving (AV), and embodied AI. A world model is an internal simulator: given current observations and a candidate action, it predicts future states of the environment—objects, agents, physics, and semantics—without executing the action in the real world.

The concept traces to model-based reinforcement learning and video prediction (e.g., Dreamer, GAIA-1). After OpenAI’s Sora (2024) demonstrated that diffusion transformers can generate coherent spatiotemporal scenes, industry focus shifted from “generate pretty video” to generate actionable futures for planning and control.

Key terms explained:

Term Meaning
World Model A learned model of environment dynamics: (s_{t+1} = f(s_t, a_t)) or latent equivalents
Latent dynamics Predictions in compressed representation space rather than raw pixels
Imagination rollouts Monte Carlo or gradient-based planning by simulating many futures in the model
Tokenization Discretizing video/state into tokens (e.g., Cosmos-Tokenizer) for transformer training

中文

2025 年初,世界模型(World Model) 从学术概念跃升为 机器人自动驾驶(AV)具身智能 的工程主线。世界模型是一种「内部模拟器」:给定当前观测与候选动作,它在不真正执行动作的前提下,预测环境的 未来状态——物体、智能体、物理规律与语义关系。

该概念源于基于模型的强化学习与视频预测(Dreamer、GAIA-1 等)。OpenAI Sora(2024)证明扩散 Transformer 能生成连贯的时空场景后,产业焦点从「生成好看视频」转向 生成可用于规划与控制的未来

关键词解释:

术语 含义
世界模型 对环境动力学的学习:(s_{t+1} = f(s_t, a_t)) 或其潜空间形式
潜空间动力学 在压缩表示而非原始像素上预测未来
想象 rollout 在模型中模拟多条未来轨迹以做规划
Tokenizer 将视频/状态离散为 token(如 Cosmos-Tokenizer),供 Transformer 训练

2025 年 1 月,NVIDIA 发布 Cosmos 世界基础模型系列,Google DeepMind 强化 Genie 2 交互式世界,Waymo 与 Tesla 在公开演讲中均将「预测 + 规划」列为 L4 核心路径。世界模型与 2024 年的 Sora 形成承接:从 生成仿真与决策

Why 2025, not earlier? Three conditions aligned: (1) petabyte-scale driving video from commercial fleets became legally usable for training; (2) diffusion transformers stabilized long clips beyond 5 seconds; (3) differentiable simulators (Genesis, Isaac) could export trajectories that match neural predictions for finetuning. Without any one pillar, world models remained academic curiosities.

与预测模块的区别: 传统 AV prediction 只 forecast 其他 agent 轨迹;世界模型 additionally 建模 ego 动作后果传感器观测演化,可反向用于 数据增强 与 ** counterfactual 测试**(「若晚 0.5 秒刹车会怎样」)。

中文补充: 对机器人而言,世界模型缓解 稀疏奖励——在想象中完成 thousands of grasp attempts 再选最优,比纯 RL 摸黑探索 sample-efficient 一个数量级。2025 产业报告普遍将 world model 与 VLA(Vision-Language-Action) 并列为具身智能双核。


二、架构 | Architecture

English

Typical 2025 world-model stacks share a perception → tokenization → dynamics → planner pipeline:

1
2
3
4
5
6
7
8
9
10
11
Sensors (camera / LiDAR / proprioception)

Encoder (ViT, BEV, or multimodal fusion)

Tokenizer / Latent compressor (Cosmos-Tokenizer, VQ-VAE)

Dynamics model (DiT, autoregressive, or JEPA-style)

Action-conditioned rollouts → cost / reward → MPC or policy

Low-level controller (trajectory, joint torques)

Architectural variants:

  1. Video-centric (Sora lineage): Diffusion Transformer predicts pixel or token sequences; actions injected via cross-attention or conditioning tokens. Strong for data-rich AV fleets; heavy compute at inference.

  2. Latent JEPA / predictive coding: Predict in representation space (I-JEPA, V-JEPA 2); lighter inference, better for real-time robotics.

  3. Physics-augmented simulators: Hybrid of neural dynamics + differentiable physics (Genesis engine) for contact-rich manipulation.

  4. Multi-scale hierarchy: High-level semantic world model (objects, relations) + low-level motion model (flows, occupancy).

中文

2025 年世界模型栈普遍遵循 感知 → Token 化 → 动力学 → 规划器 流水线:

1
2
3
4
5
6
7
8
9
10
11
传感器(相机 / 激光雷达 / 本体感知)

编码器(ViT、BEV 或多模态融合)

Tokenizer / 潜空间压缩(Cosmos-Tokenizer、VQ-VAE)

动力学模型(DiT、自回归或 JEPA 风格)

动作条件 rollout → 代价 / 奖励 → MPC 或策略网络

底层控制器(轨迹、关节力矩)

架构变体:

  1. 视频中心(Sora 谱系): 扩散 Transformer 预测像素或 token 序列;动作经 cross-attention 或条件 token 注入。适合数据丰富的 AV 车队;推理算力高。

  2. 潜空间 JEPA / 预测编码: 在表示空间预测(I-JEPA、V-JEPA 2);推理更轻,适合实时机器人。

  3. 物理增强仿真: 神经动力学 + 可微物理(Genesis)混合,适合接触丰富的操作任务。

  4. 多尺度层次: 高层语义世界模型(物体、关系)+ 低层运动模型(光流、占据栅格)。

Cosmos-Tokenizer 将连续视频压缩为离散 token,使世界模型训练与 LLM 预训练范式对齐;Genesis 提供 GPU 加速的可微物理仿真,作为世界模型的「硬约束层」或数据引擎。


English

Trend Description
From generative video to closed-loop planning Models trained not only on passive video but on action-labeled trajectories (steering, gripper commands)
Foundation world models Pretrain on internet-scale video + sim data; finetune per robot or AV domain
Long-horizon consistency 10–60 second coherent rollouts for highway merge and warehouse navigation
Safety via ensemble imagination Multiple sampled futures; risk = disagreement or collision probability across rollouts
Edge distillation Compress world models to run on vehicle NPUs alongside perception stacks

中文

趋势 说明
从生成视频到闭环规划 不仅用被动视频训练, increasingly 使用 带动作标签 的轨迹(方向盘、夹爪指令)
基础世界模型 互联网级视频 + 仿真数据预训练,再按机器人 / AV 领域微调
长时域一致性 10–60 秒连贯 rollout,支撑高速汇入、仓储导航
想象集成保安全 采样多条未来;风险 = rollout 间分歧或碰撞概率
端侧蒸馏 压缩世界模型,与感知栈一起在车载 NPU 上运行

2025 年 Q1,产业共识是:纯 端到端驾驶 与纯 模块化栈 正在融合——世界模型充当 可学习的中间仿真层,既保留数据驱动灵活性,又提供可解释的「如果这样做会发生什么」。


四、优缺点 | Pros/Cons

English

Pros

  • Sample-efficient planning: Explore millions of futures in GPU memory before one real-world move
  • Unified representation: Same model for prediction, anomaly detection, and synthetic data generation
  • Transfer from video pretraining: Leverage Sora-class models and web video for priors on physics and scenes
  • Interpretability (relative): Visualized rollouts help engineers debug failure modes

Cons

  • Compounding error: Long rollouts drift; small prediction errors accumulate into unsafe plans
  • Sim-to-real gap: Learned dynamics may fail on rare objects, weather, or contact physics
  • Compute cost: Full diffusion rollouts are too slow for 10 Hz control without distillation
  • Evaluation difficulty: No single metric captures “world understanding”; benchmarks still immature

中文

优点

  • 样本高效规划: 在 GPU 记忆中探索百万条未来,再执行一次真实动作
  • 统一表示: 同一模型可用于预测、异常检测与合成数据生成
  • 视频预训练迁移: 利用 Sora 级模型与网络视频获得物理与场景先验
  • 相对可解释: 可视化 rollout 便于工程师调试失效模式

缺点

  • 误差累积: 长 rollout 漂移;小预测误差叠加为不安全规划
  • 仿真到真实鸿沟: 罕见物体、天气、接触物理上动力学可能失效
  • 算力成本: 完整扩散 rollout 难以支撑 10 Hz 控制,需蒸馏
  • 评估困难: 尚无单一指标衡量「世界理解」;基准仍不成熟

五、应用场景 | Use Cases

English

Domain Use case World model role
Highway AV Lane change, merge Predict surrounding vehicle reactions; score trajectories
Urban robotaxi Jaywalker, construction zones Semantic future occupancy; conservative planning
Warehouse AMR Forklift + human coexistence Short-horizon interaction prediction
Manipulation Pick-place in clutter Imagine object motion after push or grasp
Humanoid Locomotion on uneven terrain Contact-rich rollout in Genesis-class sim
Training data Rare event synthesis Generate corner-case video for perception finetuning

中文

领域 场景 世界模型作用
高速 AV 变道、汇入 预测周围车辆反应;轨迹打分
城市 Robotaxi 行人横穿、施工区 语义未来占据;保守规划
仓储 AMR 叉车与人共存 短 horizon 交互预测
操作臂 clutter 中取放 想象推/抓后物体运动
人形机器人 uneven 地形行走 Genesis 类仿真中接触丰富 rollout
训练数据 罕见事件合成 生成 corner case 视频供感知微调

六、GitHub 开源生态 | GitHub

English

Repository Relevance
openai/sora Reference architecture for spatiotemporal diffusion; inspires world-model tokenization and scaling laws
NVIDIA/Cosmos-Tokenizer Efficient video tokenization for world foundation model pretraining
genesis-embodied-ai/Genesis Universal physics engine for embodied AI; pairs with learned world models for sim2real

中文

仓库 关联
openai/sora 时空扩散参考架构;影响世界模型 token 化与缩放定律
NVIDIA/Cosmos-Tokenizer 高效视频 Token 化,支撑世界基础模型预训练
genesis-embodied-ai/Genesis 具身 AI 通用物理引擎;与学习型世界模型配合做 sim2real

七、参考资料 | References

English

  1. OpenAI — Sora: Video generation models as world simulators (2024–2025 technical reports)
  2. NVIDIA — Cosmos World Foundation Models announcement (CES 2025)
  3. LeCun et al. — A path towards autonomous machine intelligence (JEPA framework)
  4. Ha & Schmidhuber — World Models (original RL formulation)
  5. Waymo Research — Scene Flow and prediction stack updates (2025)

中文

  1. OpenAI — Sora:作为世界模拟器的视频生成模型(2024–2025 技术报告)
  2. NVIDIA — Cosmos 世界基础模型发布(CES 2025)
  3. LeCun 等 — 通向自主机器智能之路(JEPA 框架)
  4. Ha & Schmidhuber — World Models(原始 RL 表述)
  5. Waymo Research — 场景流与预测栈更新(2025)

八、工程落地清单 | Engineering Checklist

English

Before deploying a world model in production AV or robotics (2025 vendor checklists):

  1. Horizon budget: Define prediction seconds vs. control frequency; match model output rate to planner needs (often 2–10 Hz semantic, 50 Hz low-level).
  2. Uncertainty calibration: Ensemble or dropout-based epistemic estimates; block actions when variance exceeds threshold.
  3. Real-world anchoring: Periodic sync with live perception—prevent imagination drift over multi-step plans.
  4. Regression corpus: Store failed rollouts from field incidents; retrain on counterfactual labels.
  5. Compute partition: Train on GPU cluster; distill to TensorRT / NPU for onboard short-horizon module.

中文

生产 AV 或机器人部署世界模型前(2025 厂商清单):

  1. Horizon 预算: 明确预测秒数 vs 控制频率;模型输出率匹配规划器(常 2–10 Hz 语义、50 Hz 底层)。
  2. 不确定性校准: 集成或 dropout 认知不确定性;超阈则 block 动作。
  3. 真实锚定: 定期与 live 感知 sync——防多步规划想象漂移。
  4. 回归语料: 存储现场事故失败 rollout;用 counterfactual 标签再训。
  5. 算力分区: 集群训练;蒸馏至 TensorRT/NPU 作 onboard 短 horizon 模块。

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 年的世界模型是连接 AIGC 视频能力物理世界决策 的桥梁。其工程价值在于「在想象中试错」;落地瓶颈仍是长 horizon 精度、实时推理与可验证安全。与 Sora、Cosmos、Genesis 的开源与半开源生态共同构成本年度机器人 / AV 算法主线。

English: In 2025, world models bridge generative video and physical decision-making. Their engineering value is trial-and-error in imagination; deployment bottlenecks remain long-horizon accuracy, real-time inference, and verifiable safety. Together with Sora, Cosmos, and Genesis ecosystems, they define the year’s robotics and AV algorithm stack.