AI 技术编年史 2024：具身智能与人形机器人

发表于 分类于 阅读次数：
2024 年具身智能与人形机器人商业化加速：VLA 模型、Figure、Optimus、宇树与智元的中英文产业解读。

具身智能与人形机器人 | Embodied AI and Humanoid Robots

一、背景与核心概念 | Background and Core Concepts

English

Embodied AI refers to intelligence grounded in physical interaction — robots, vehicles, and devices that perceive, act, and learn in the real world. In 2024, humanoid robots became the public face of embodied AI, driven by advances in Vision-Language-Action (VLA) models, sim-to-real transfer, and LLM-based task planning.

Key concepts:

  • VLA models: unified models mapping camera input + language instructions → motor actions (e.g., RT-2, OpenVLA)
  • Teleoperation + imitation learning: human demos bootstrap policy learning
  • Sim-to-real: Isaac Sim, MuJoCo, and domain randomization bridge simulation gaps
  • Foundation models for robotics: LLMs decompose “clean the kitchen” into sub-skills

2024 milestones included Figure 01 + OpenAI, Tesla Optimus factory trials, Unitree H1/G1 price drops, and Agibot (智元) mass-production announcements in China.

中文

具身智能指扎根于物理交互的智能——机器人、车辆、设备在真实世界中感知、行动与学习。2024 年人形机器人成为具身智能公众符号，驱动力来自 VLA（视觉-语言-动作） 模型、Sim-to-Real 与 LLM 任务规划。

核心概念：VLA 统一模型（RT-2、OpenVLA）；遥操作 + 模仿学习；Isaac Sim 等仿真迁移；LLM 将高层指令分解为子技能。

2024 里程碑：Figure 01 接入 OpenAI、Tesla Optimus 进工厂、宇树 H1/G1 降价、智元量产发布等。

术语 含义
End-effector 末端执行器（手、夹爪）
Biped locomotion 双足行走平衡控制
Dexterous manipulation 灵巧手精细操作
Embodied cognition 认知依赖身体与环境的理论

1.1 从深度学习到 VLA | From Deep Learning to VLA

English

Classical robotics used modular stacks (SLAM → planning → control) with hand-tuned parameters. 2024’s VLA models unify perception and action in one network, pretrained on internet-scale vision-language data then fine-tuned on robot trajectories. Google DeepMind’s RT-2 showed transferring web knowledge (“pick up the extinct animal” → dinosaur toy) — a capability impossible with narrow imitation-only policies.

Humanoid form factor returned not from engineering necessity but environment fit: warehouses and homes built for bipedal humans; dual arms enable tool use. Tesla, Figure, and Chinese vendors bet humanoids over wheeled arms for general-purpose branding.

中文

经典机器人用模块化栈（SLAM→规划→控制）与手工调参。2024 VLA 在单网统一感知与动作，先 Web 规模 VLM 预训练再机器人轨迹微调。Google RT-2 展示 Web 知识迁移（「捡起灭绝动物」→恐龙玩具）——窄模仿策略无法实现。人形回归非工程必然，而是环境适配：人机共存空间为双足设计；双臂支持工具。特斯拉、Figure 与中国厂商押注人形而非轮式臂，求通用品牌叙事。

二、架构设计 | Architecture

English

A modern humanoid stack layers perception, cognition, and control:

1
2
3
4
5
6
7
8
9
10
11
Sensors (RGB-D, IMU, Force/Torque, Proprioception)
    ↓
Perception Module (detection, segmentation, SLAM)
    ↓
High-level Planner (LLM / VLM task decomposition)
    ↓
Mid-level Policy (VLA / diffusion policy / RL)
    ↓
Low-level Controller (whole-body MPC, PID, impedance)
    ↓
Actuators (harmonic drives, linear actuators, dexterous hands)

Training loop: collect teleop data → train policy in sim → fine-tune on real robot → continuous learning from failures.

中文

现代人形栈分层：传感器 → 感知 → 高层规划（LLM/VLM）→ 中层策略（VLA/扩散策略/RL）→ 底层控制（MPC、阻抗）→ 执行器。训练闭环：遥操作采集 → 仿真预训练 → 真机微调 → 失败持续学习。

2.1 VLA 模型架构 | VLA Model Architecture

组件 功能
Vision Encoder ViT / SigLIP 处理多帧图像
Language Encoder 指令语义理解
Action Head 输出关节角、末端位姿或 action tokens
History Buffer 短时序上下文

RT-2 范式：将机器人动作离散化为 token，与 Web 规模 VLM 联合 co-fine-tune，实现 zero-shot 泛化到新物体。

2.2 2024 代表性平台 | 2024 Representative Platforms

平台 亮点
Figure 01 OpenAI 多模态模型驱动对话 + 操作
Tesla Optimus 工厂内搬运试点，FSD 团队 overlap
Unitree G1 约 9.9 万元级定价，开发者友好
智元 Agibot 中国量产叙事，远征 A1/A2
1X NEO 家用场景，软体安全设计

English: None achieved millions of units shipped in 2024 — all remained pilot / demo / early order phase. The year was about proof of learning curves, not mass replacement of factory workers.

中文：2024 无一实现百万台出货——均为试点/演示/早期订单。当年重在学习曲线证明，非大规模替代产线工人。

English

2024 embodied AI trends:

  1. Humanoid hype cycle — billions in funding (Figure, 1X, Sanctuary, Chinese startups)
  2. LLM as robot brain — natural language tasking replaces rigid state machines
  3. Cost reduction — Unitree G1 under $16k signaled consumer-adjacent pricing
  4. Manufacturing pilots — BMW, Tesla test humanoids for repetitive tasks
  5. Data scarcity — robot data orders of magnitude smaller than web text; synthetic data rising
  6. Safety regulation — EU AI Act and workplace safety standards emerging

中文

2024 趋势：人形机器人融资热潮；LLM 作”机器人大脑”；宇树 G1 低价信号；宝马、特斯拉工厂试点；机器人数据稀缺推动合成数据；欧盟 AI 法案与安全标准酝酿。

四、优缺点分析 | Pros and Cons

4.1 优点 | Advantages

  1. 通用形态 — 人形适配人类环境 / Human-shaped for human environments
  2. 多模态融合 — VLA 统一感知-语言-动作 / Unified VLA pipelines
  3. LLM 规划 — 自然语言任务接口 / Natural language tasking
  4. 仿真加速 — 降低真机试错成本 / Sim reduces real-world trial cost
  5. 劳动力补充 — 危险/重复岗位自动化 / Automation for dull/dangerous jobs
  6. 研究汇聚 — CV + NLP + 控制交叉创新 / Cross-disciplinary innovation

4.2 缺点 | Disadvantages

  1. 可靠性不足 — 2024 真机仍易失败 / Fragile real-world performance
  2. 成本与维护 — 硬件、维修、能耗高 / High TCO
  3. 数据瓶颈 — 缺乏 Web 级机器人数据 / Robot data scarcity
  4. 安全风险 — 物理伤害与 liability / Physical safety concerns
  5. 泛化有限 — 新环境迁移仍难 / Poor out-of-distribution generalization
  6. 炒作 vs 现实 — 演示与量产差距大 / Demo-to-production gap

五、典型应用场景 | Use Cases

场景 Scenario 中文说明 English Description
工厂搬运 产线物料、货架拣选 Material handling on factory floors
物流仓储 与 AGV 协同的灵巧操作 Warehouse picking with mobility
家庭服务 整理、简单清洁（早期试点） Home tidying and assistive tasks
养老护理 辅助起身、物品递送 Elder care assistance
灾难救援 危险环境侦察与操作 Hazardous environment operations
科研平台 算法验证与数据采集 Research testbeds for VLA

六、GitHub 与开源生态 | GitHub and Open Source

English

Open robotics ecosystem in 2024:

  • openvla/openvla: open VLA model weights and training code
  • google-deepmind/open_x_embodiment: large-scale multi-robot dataset
  • facebookresearch/habitat: embodied AI simulation
  • unitreerobotics: Unitree SDK and sim tools
  • ros2: Robot Operating System middleware

中文

开源生态：OpenVLA 模型、Open X-Embodiment 数据集、Habitat 仿真、宇树 SDK、ROS 2 中间件。

仓库 说明
openvla/openvla 开源 VLA 模型
google-deepmind/open_x_embodiment 跨本体机器人数据
facebookresearch/habitat-lab 具身 AI 仿真
ros2/ros2 机器人操作系统

七、参考链接 | References

  • RT-2: Vision-Language-Action Models (Google DeepMind)
  • OpenVLA 论文与模型卡
  • Figure AI + OpenAI 合作公告
  • Tesla Optimus 2024 进展更新
  • 宇树 H1/G1 产品页：unitree.com
  • Isaac Sim 文档：developer.nvidia.com/isaac/sim

八、2025 展望 | Outlook for 2025

English

2025–2026 roadmaps point to VLA models in consumer robots (sub-$20k humanoids), simulation-generated data exceeding human teleop, and LLM planners with verifiable skill libraries. Regulatory frameworks for workplace humanoids (EU Machinery Regulation updates) will clarify liability. Investment may consolidate after 2024 hype — survivors combine hardware margin + software subscription (RaaS). Technical moat shifts from walking demos to reliability metrics: MTBF, success rate on 1000-task benchmarks, OTA improvement velocity.

中文

2025–2026 路线：VLA 进消费级机器人（2 万美元以下人形）、仿真数据超 human teleop、LLM 规划+可验证技能库。 workplace 人形法规（欧盟机械法规更新）将厘清责任。投资或在 2024 hype 后整合——幸存者靠硬件毛利+软件订阅（RaaS）。技术护城河从行走 demo 转向可靠性指标：MTBF、千任务成功率、OTA 改进速度。

English Summary: 2024 embodied AI moved from lab demos toward factory pilots — VLA models and LLM planners converged, but reliability and data remain the gating factors for commercial scale.

中文总结：2024 具身智能从实验室演示走向工厂试点——VLA 与 LLM 规划汇聚，可靠性与数据仍是规模化门槛。