Unitree Scales $6,000 Humanoids as Physical AGI Faces Five-Year Delay

Insights for the Humanoid and Quadruped Robot Industry With Unitree Robotics — Ai4

Ai4

Disembodied AI foundation models cannot achieve physical autonomy without unified world models and direct sensorimotor interaction, pushing unstructured general-purpose deployments beyond 2030.

AI-assisted summary of Insights for the Humanoid and Quadruped Robot Industry With Unitree Robotics — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 29, 2025, this presentation features Tony Yang, North American Director at Unitree Robotics. The session discusses how recent breakthroughs in artificial intelligence, physical computing, and foundation models are enabling embodied AI and driving the rapid commercialization of quadruped and humanoid robots.

## TL;DR Unitree Robotics outlines the three-stage evolution of humanoid robotics, tracing the shift from early lab-constrained prototypes to generative AI-powered embodied agents and future physical AGI. The talk details Unitree's hardware trajectory from cost-efficient quadrupeds to the H1, G1, and sub-$6,000 R1 humanoid platforms driven by vertically integrated manufacturing. It underscores the limitations of disembodied AI, arguing that genuine physical intelligence requires sensorimotor embodiment rather than passive video learning. Finally, it explores reinforcement learning sim-to-real pipelines and highlights the need for unified world models to bridge the gap toward versatile, real-world autonomy.

## Atmosphere & style The presentation maintains a pragmatic, fast-paced, and highly technical tone that cuts through broad AI hype with grounded engineering realities. Yang speaks with direct industry clarity, balancing the tangible mechanics of joint torque, cost reduction, and sim-to-real gaps with an ambitious vision for physical AGI. Listening to the talk feels like an insider briefing on rapid hardware-software co-design, pairing academic benchmarks with the relentless velocity of commercial robotics manufacturing.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) The evolution of humanoid robotics is categorized into three eras: the pre-2021 lab hardware era, the 2021–2030 generative AI integration wave, and the post-2030 physical AGI era. - [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) Post-2030 physical AGI is projected to deploy sequentially, beginning in structured industrial environments like factories before moving into unstructured domestic households. - [02:11](https://www.youtube.com/watch?v=BMEsifCJL_c&t=131s) Unitree's rapid iteration and supply chain cost reductions stem from vertical integration, manufacturing its own electric actuators, robotic arms, and LiDAR units in-house. - [02:11](https://www.youtube.com/watch?v=BMEsifCJL_c&t=131s) After breaking the sub-$10,000 quadruped cost barrier in 2020, Unitree expanded into consumer robot dogs and industrial wheeled-leg platforms for inspection and emergency rescue. - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) Unitree greenlit its bipedal humanoid program in late 2022 after generative AI breakthroughs, developing the full-size H1 prototype in roughly six months with a two-engineer team. - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) The full-size H1 demonstrated the world's first electric humanoid standing backflip, while the compact G1 platform achieved a 1.4-meter dynamic vertical jump. - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) Accelerated physics simulation engines like Genesis, Isaac Lab, and Isaac Gym have compressed humanoid locomotion iteration cycles from months to weeks. - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) Mass-manufacturing efficiencies and scaled actuator production enabled Unitree to introduce the low-cost R1 humanoid platform at approximately $6,000. - [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) Academic institutions have demonstrated complex autonomous recovery, dynamic martial arts, video mimicry, and multi-agent coordination on Unitree hardware. - [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) Disembodied LLMs cannot achieve physical agency without embodied hardware and physical interaction, just as humans cannot learn to ride a bicycle solely by watching videos. - [14:22](https://www.youtube.com/watch?v=BMEsifCJL_c&t=862s) Coordinating a humanoid robot with up to 60 degrees of freedom in real time is an order of magnitude more complex than autonomous driving control systems. - [14:22](https://www.youtube.com/watch?v=BMEsifCJL_c&t=862s) Current robotics foundation models face key deficits: lack of autonomous self-learning, poor knowledge transfer between skills, and fragmented paradigms between navigation and manipulation. - [14:22](https://www.youtube.com/watch?v=BMEsifCJL_c&t=862s) Physical AGI requires unified world models that view both manipulation and navigation as identical end-effector interactions learning through continuous self-supervised exploration. - [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) Modern dynamic motion synthesis relies on capturing human motion data, optimizing reinforcement learning policies within physics simulators, and transferring weights zero-shot to physical robots. - [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) Production-ready robotic foundation world models are estimated to be 3 to 5 years away, with early systems like Google Gemini Robotics already demonstrating direct video-to-action control.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) The Three Stages of Humanoid Robotics Evolution

Unitree Robotics frames the trajectory of humanoid robotics across three distinct evolutionary eras spanning decades. The foundational early period, taking place prior to 2021, established the mechanical groundwork with landmark platforms such as Boston Dynamics' Spot and Honda's ASIMO. While these pioneering machines sparked widespread academic and corporate interest across the United States and China, the sector remained heavily constrained. Only a negligible number of companies possessed the technical and industrial capability for mass manufacturing, and the systems operated primarily as expensive laboratory research testbeds hampered by severe hardware bottlenecks rather than serving viable commercial scenarios.

The arrival of ChatGPT and breakthroughs in generative AI around late 2022 fundamentally shifted the landscape, initiating the second wave spanning 2021 to 2030. Generative AI catalyzed unprecedented venture capital and industrial investment into robotics. Large language models and emerging Artificial General Intelligence (AGI) concepts are currently injecting high-level reasoning into physical embodiments, drawing universities and enterprises into an accelerating research race. Looking beyond 2030, the third era will mark the arrival of "real AGI" in physical systems, which Unitree anticipates will unfold across two sequential deployment phases: structured industrial environments such as factories and manufacturing plants first, followed by complex, unstructured environments including private households.

```mermaid flowchart LR accTitle: Three Stages of Humanoid Robotics subgraph Early["Pre-2021: Early Lab Era"] A["Pioneering Hardware<br/>(Spot, ASIMO)"] --> B["Academic Research &<br/>Hardware Bottlenecks"] end subgraph Transition["2021–2030: Generative AI Boom"] C["ChatGPT & Foundation Models"] --> D["Capital Influx & Embodied AI Integration"] end subgraph Future["Post-2030: Physical AGI"] E["Structured Industrial Scenarios<br/>(Manufacturing & Logistics)"] --> F["Unstructured Scenarios<br/>(Home & Domestic Use)"] end Early --> Transition Transition --> Future ```

- **Pre-2021 Era**: Defined by high-cost hardware platforms, bespoke R&D, and negligible commercial scalability. - **2021–2030 Horizon**: Massive ecosystem expansion driven by generative AI, foundation model integration, and rapid simulation-to-real locomotion advances. - **Post-2030 Physical AGI**: Transition from constrained industrial manufacturing to fully unstructured, home-use embodied agents.

---

[02:11](https://www.youtube.com/watch?v=BMEsifCJL_c&t=131s) Unitree's Hardware Journey: From Quadrupeds to Wheeled-Leg Robots

Unitree's hardware foundation traces back to the master's research of its founder, Xingxing Wang, who developed the prototype XDog quadruped between 2013 and 2016. Upon graduating in 2016, Wang formally established Unitree Robotics to commercialize legged systems. The company pursued an aggressive vertically integrated component strategy, designing and manufacturing its own high-performance electric actuators, robotic arms, and custom LiDAR units in-house. Selling core actuator modules independently while integrating them into proprietary robots allowed Unitree to drive rapid iteration cycles and achieve significant supply-chain cost reductions across successive hardware generations.

The rollout of the Aliengo quadruped in 2019 delivered a much more compact form factor, paving the way for the breakthrough in 2020 that brought quadruped production costs under $10,000. By 2021, Unitree pioneered consumer-grade quadrupedal robotics, subsequently launching the Go2 quadruped platform in 2023. Alongside its consumer line, Unitree expanded into industrial automation with the heavy-duty B2 quadruped and specialized wheeled-leg variants (Go2-W and B2-W), targeting demanding inspection, firefighting, and emergency rescue operations.

| Platform / Milestone | Timeline | Key Characteristics & Significance | |:--- |:--- |:--- | | **XDog Prototype** | 2013–2016 | Founder Xingxing Wang's master's thesis quadruped research. | | **Unitree Founded** | 2016 | Official corporate establishment and first-generation quadruped R&D. | | **Aliengo** | 2019 | Compact, modular second-generation commercial quadruped. | | **Sub-$10,000 Milestone** | 2020 | Aggressive bill-of-materials reduction through in-house actuator development. | | **Consumer Quadruped** | 2021 | World's first mass-market consumer-grade robot dog. | | **Go2 & Wheeled-Leg Lines** | 2023 | Go2 release alongside Go2-W and industrial B2/B2-W wheeled-leg systems. |

---

[04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) Humanoid Lineup: H1, G1 Agility, and the $6,000 R1

Before late 2022, Unitree intentionally stayed away from bipedal humanoid development because disembodied AI algorithms lacked the motor intelligence required for general utility. However, the acceleration of generative AI led Unitree to greenlight its humanoid program at the close of 2022. Working with remarkable speed, a team of just two lead engineers constructed the functional prototype of the full-size H1 humanoid robot in roughly six months. Continuous software updates followed: by early 2024, the H1 executed the world's first standing backflip achieved by a full-sized, fully electrically powered humanoid robot, accompanied by an open-source teleoperation framework released to developers.

In May 2024, Unitree unveiled the G1, a compact, highly agile humanoid standing approximately 1.3 meters tall, weighing 35 kg, equipped with optional three-finger dexterous hands, and generating peak joint torques of up to 120 N·m. By October 2024, locomotion updates pushed the hardware beyond initial design parameters, enabling the G1 to achieve dynamic vertical jumps reaching 1.4 meters. Accelerated modern physics engines—including Genesis, Isaac Lab, and Isaac Gym—have compressed iteration cycles to mere weeks by accurately modeling friction, gravity, and contact dynamics. Expanding the portfolio further, Unitree introduced the R1 humanoid, priced at roughly $6,000, leveraging mature mass-manufacturing pipelines and scaled actuator production.

```mermaid flowchart TD accTitle: Unitree Humanoid Product Progression subgraph H1Progression["Full-Size H1 Platform"] H1A["Late 2022: Project Greenlit"] --> H1B["Mid-2023: Prototype Built in 6 Months"] H1B --> H1C["Early 2024: World's First Full-Size Electric Backflip"] H1C --> H1D["Aug 2024: Open-Source Teleoperation Stack"] end subgraph G1Progression["Compact Agile G1 Platform"] G1A["May 2024: G1 Released (1.3m, 35kg, 120 Nm)"] --> G1B["Oct 2024: 1.4m Dynamic Jump & Combat Testing"] G1B --> G1C["Early 2025: Rapid Bi-Monthly Locomotion Policy Refinements"] end subgraph R1Release["Mass-Market R1 Platform"] R1A["~ $6,000 Price Point via Scaled Mass Production"] end H1Progression --> G1Progression G1Progression --> R1Release ```

- **H1 Full-Size Humanoid**: Built in six months by a two-engineer core team; demonstrated the first electric full-size humanoid backflip. - **G1 Specifications**: 1.3 m height, 35 kg weight, 120 N·m maximum joint torque, optional 3-finger dexterous hands, and a 1.4 m dynamic jump capability. - **R1 Low-Cost Humanoid**: Priced at approximately $6,000 to lower the barrier for robotics researchers by transferring G1 mass-production economies of scale to a slightly smaller kinematic frame.

---

[08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) Academic Breakthroughs and the Limits of Disembodied AI

A surge of academic benchmarks illustrates the rapid expansion of embodied locomotion and manipulation research on Unitree platforms: - **UIUC (University of Illinois Urbana-Champaign)**: Developed omnidirectional recovery policies allowing humanoids to stand up autonomously from any arbitrary fallen posture. - **UC Berkeley & Tsinghua University**: Trained dynamic martial arts and Bruce Lee motion-imitation behaviors. - **UC San Diego (UCSD)**: Deployed the X-Body 2 framework enabling expressive bipedal dancing and complex teleoperation-to-autonomous hand manipulation. - **Stanford University & Berkeley**: Integrated large language models for high-level semantic execution and demonstrated "VideoMimic", where robots learn locomotion, sitting, and standing directly from raw monocular human video feeds. - **NVIDIA & CMU**: Trained multi-agent dynamic coordination and basketball playing skills.

Despite these lab demonstrations, disembodied intelligence faces fundamental limitations. A disembodied Large Language Model attempting physical control is akin to "falling down in a dream" where "you do not have a physical body" and "you always cannot feel that you can run fast." Neural capacity alone does not guarantee physical agency: an octopus possesses over 500 million neurons, parrots possess 200 million to 3 billion neurons, and humans possess 16 billion cortical neurons. Yet parrots cannot do laundry or cook because they lack hands, and monkeys cannot articulate human speech due to vocal tract constraints. Just as a human cannot learn to ride a bicycle or drive a car purely by passively watching videos without hands-on practice, robots cannot achieve 100% real-world success rates without direct physical interaction and reinforcement learning on real hardware.

---

[14:22](https://www.youtube.com/watch?v=BMEsifCJL_c&t=862s) Hardware Complexity, Unified Architectures, and the Path to World Models

Controlling a humanoid robot presents an order-of-magnitude leap in complexity compared to autonomous vehicles. Autonomous driving systems typically map simple sensor inputs to narrow steering, acceleration, and braking control outputs. In contrast, a humanoid robot with five-finger dexterous hands possesses up to 60 independent joints and degrees of freedom that must be coordinated simultaneously in real time. The robotics industry continues to debate whether proprioceptive control or dense external sensor feedback offers the superior control foundation.

```mermaid flowchart TD accTitle: Embodied AI Architectural Challenges subgraph Bottlenecks["Core Algorithmic Deficits"] N1["Data-Hungry Foundation Models<br/>(Lacks Autonomous Self-Learning)"] N2["Sim-to-Real Reality Gap &<br/>Reinforcement Learning Sample Inefficiency"] N3["Fragmented Task Models<br/>(No Knowledge Transfer between Kung Fu & Dance)"] end subgraph Solution["Unified World Model Vision"] U1["Unified Interaction Representation<br/>(Dexterous Manipulation ≡ Legged Navigation)"] U2["Continuous Self-Supervised Experience Accumulation<br/>(Child-like Intuitive Physics Learning)"] U3["Direct Real-Time Physical Embodiment Integration"] end Bottlenecks --> Solution ```

Current AI foundation models suffer from three structural deficits in physical settings: 1. **Lack of Autonomous Self-Learning**: Existing models rely strictly on fixed training corpora rather than self-supervised physical exploration, making them brittle when generalizing to novel dynamics. 2. **Poor Sample Efficiency and Knowledge Retention**: Reinforcement learning policies require massive simulated rollouts, yet mastering one skill (such as dancing) provides zero transferrable physical intuition when learning another skill (such as martial arts). 3. **Fragmented Architectural Paradigms**: Manipulation and navigation are currently modeled as distinct problems, whereas they are fundamentally identical: both involve an embodied end-effector interacting dynamically with the physical environment.

Achieving physical AGI demands unified "World Models" that continuously accumulate real-time sensorimotor experience—mirroring a child who learns to grasp a cup through intuitive trial and error rather than explicit instruction.

---

[17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) Q&A: Motion Capture Sim-to-Real Pipelines and Google Gemini World Models

During the closing question-and-answer session, attendees pressed on the exact technical pipeline for training dynamic motor skills and the realistic timeline for functional world models in production robotics.

```mermaid flowchart LR accTitle: Modern Humanoid Sim-to-Real Pipeline A["1. Human Motion Capture<br/>(MoCap / Video Demonstration)"] --> B["2. High-Fidelity Simulation<br/>(Isaac Gym / Isaac Lab)"] B --> C["3. Policy Optimization<br/>(Reinforcement Learning)"] C --> D["4. Sim-to-Real Transfer<br/>(Zero-Shot Physical Deployment)"] ```

Explaining the motion synthesis workflow, Unitree details that high-dynamic behaviors (such as martial arts or complex dancing) begin with motion capture systems recording human movement. These kinematic trajectories are ingested into accelerated physics simulators like Isaac Lab and Isaac Gym, where reinforcement learning policies are trained against simulated gravity, contact, and inertia. Once policy convergence is achieved, the resulting network weights are deployed directly onto the physical robot via sim-to-real transfer frameworks.

When asked about the timeline for true foundation world models in robotics, the assessment is that production-ready deployment is still roughly three to five years away. Nevertheless, early precursors are already demonstrating promising direct-action synthesis. Google's Gemini Robotics research exemplifies this trajectory by consuming multi-modal video feeds and directly outputting low-level joint control trajectories to manipulate physical objects without requiring fragmented intermediate planners.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=BMEsifCJL_c&t=0s) Unitree have been working for more than 8 years in leg robot and 1 year in humanoid robot area. — Tony Yang

> [00:00](https://www.youtube.com/watch?v=BMEsifCJL_c&t=0s) There are three most important parts, algorithm, data and computing capability for realized AGI. — Tony Yang

> [00:00](https://www.youtube.com/watch?v=BMEsifCJL_c&t=0s) build robust physical humanoid robot system is key for this ecosystem, and World Large-Scale Model (most people called foundation model) is the key to bring Embodied AI for for humanoid robot — Tony Yang

> [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) falling down in a dream — Tony Yang

> [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) you do not have a physical body — Tony Yang

> [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) you always cannot feel that you can run fast. — Tony Yang

> [08:53](https://www.youtube.com/watch?v=BMEsifCJL_c&t=533s) VideoMimic — Tony Yang

> [14:22](https://www.youtube.com/watch?v=BMEsifCJL_c&t=862s) World Models — Tony Yang

## Mentioned Tony Yang, Xingxing Wang, Unitree Robotics, Boston Dynamics, Honda, ChatGPT, OpenAI, Spot, ASIMO, XDog, Aliengo, Go2, Go2-W, B2, B2-W, H1, G1, R1, Genesis, Isaac Lab, Isaac Gym, NVIDIA, University of Illinois Urbana-Champaign (UIUC), UC Berkeley, Tsinghua University, UC San Diego (UCSD), X-Body 2, Stanford University, Carnegie Mellon University (CMU), Google, Gemini

## Video description Tony Yang, North American Director, Unitree Robotics

Abstract

With artificial intelligence technology move very fast in the past two years, humanoid robot have been one of the most import form to realized embodied AI and AGI, Unitree have been working for more than 8 years in leg robot and 1 year in humanoid robot area. There are three most important parts, algorithm, data and computing capability for realized AGI. Those three part will finally running on physical robots, we believe build robust physical humanoid robot system is key for this ecosystem, and World Large-Scale Model (most people called foundation model) is the key to bring Embodied AI for for humanoid robot, we will share the most important progressing have been made on industry and research side in the past one year, and expect and excited for new progressing will happening in next few years soon.

[Source](https://www.youtube.com/watch?v=BMEsifCJL_c)

中文简报

宇树规模化推进6000美元人形机器人，物理AGI或面临五年推迟 在缺乏统一世界模型与直接感觉运动交互的情况下，非具身AI基础模型无法实现物理自主，导致非结构化通用场景的部署被推迟至2030年之后。 宇树科技（Unitree Robotics）概述了人形机器人的三阶段演进历程，梳理了从早期局限于实验室的原型，到生成式AI驱动的具身智能体，再到未来物理AGI的演变。演讲详细介绍了宇树在垂直一体化制造推动下的硬件发展轨迹，从高性价比的四足机器人逐步拓展至H1、G1以及低于6000美元的R1人形机器人平台。同时，内容强调了非具身AI的局限性，认为真正的物理智能需要感觉运动层面的具身交互，而非被动的视频学习。最后，演讲探讨了强化学习从仿真到真实（Sim-to-Real）的迁移管线，并指出需要统一的世界模型来弥合差距，以实现通用的现实世界自主能力。 人形机器人演进被划分为三大阶段：2021年前的实验室硬件期、2021至2030年的生成式AI深度融合期，以及2030年后预计先在工厂落地、再逐步走进家庭环境的物理AGI时代。 (0:04) 宇树科技通过自研自产电机执行器、机械臂和LiDAR等垂直整合策略，在2020年将四足机器人成本降至1万美元以下，并持续拓展至消费级机器狗与工业级轮足平台。 (2:11) 2022年底启动双足项目后，宇树在6个月内打造出实现全尺寸电驱人形后空翻的H1原型，随后推出动态跳跃达1.4米的G1平台，并依托规模量产能力发布了售价约6,000美元的R1平台。 (4:33) 当前离身AI模型在物理世界中面临缺乏自监督探索与知识泛化等瓶颈，业界正通过动捕结合Isaac等物理引擎加速Sim-to-Real迁移，预计端到端具身世界模型将在3至5年内迈向生产级应用。 (17:34)

Reading summary

Unitree Robotics argues that achieving viable physical artificial general intelligence requires moving beyond disembodied foundation models toward vertically integrated hardware capable of continuous physical interaction. While generative AI and accelerated simulation have compressed development cycles and reduced hardware costs, general utility remains constrained by algorithmic deficits that unified world models will take an estimated three to five years to solve. Three Evolutionary Eras of Robotics (0:04) Unitree frames the trajectory of humanoid robotics across three distinct phases. Prior to 2021, platforms such as Boston Dynamics' Spot and Honda's ASIMO established mechanical foundations but operated primarily as expensive, lab-bound testbeds hampered by severe hardware bottlenecks and negligible mass-manufacturing capability. The current era spanning 2021 to 2030, catalyzed by ChatGPT and breakthroughs in generative AI, has channeled substantial capital into integrating foundation models and high-level reasoning into physical embodiments. Looking beyond 2030, Unitree forecasts the emergence of true physical artificial general intelligence (AGI). The company anticipates deployment will occur in two sequential waves: first targeting structured industrial settings like manufacturing plants and logistics facilities, followed later by complex, unstructured environments such as private households. Vertical Integration and Rapid Hardware Iteration (2:11) Commercialization across legged robotics has relied heavily on aggressive vertical integration and component-level cost reductions. Tracing back to founder Xingxing Wang's master's research on the XDog prototype from 2013 to 2016, Unitree established in-house design and manufacturing for electric actuators, robotic arms, and custom LiDAR sensors. Selling actuators independently alongside integrated robots enabled rapid iteration cycles, bringing quadruped production costs below $10,000 in 2020 and enabling the launch of mass-market consumer models like the Go2 by 2021. After avoiding bipedal systems due to algorithmic limits, Unitree initiated its humanoid program in late 2022. A two-engineer core team built the full-size H1 prototype in roughly six months, later achieving the first standing backflip by a fully electric humanoid. The company subsequently introduced the compact G1—standing 1.3 meters, weighing 35 kg, with 120 N·m joint torque and a 1.4-meter vertical jump—and the R1 humanoid, priced at approximately $6,000 by transferring quadruped mass-manufacturing economies of scale to bipedal frames. Disembodied Intelligence Versus Physical Morphology (8:53) A core contention in Unitree's assessment is that disembodied large language models cannot master real-world motor execution simply through software scaling. Unitree compares an LLM attempting physical tasks to dreaming without a physical body, where the system cannot gauge real dynamic feedback. Biological comparisons highlight that neurological capacity does not equal physical capability: octopuses have over 500 million neurons and parrots possess up to 3 billion, yet neither can execute household chores because they lack human-like hands, just as monkeys cannot produce human speech due to vocal tract constraints. Furthermore, coordinating a humanoid represents an order-of-magnitude leap in complexity over autonomous vehicles. While autonomous driving systems map sensor inputs to a narrow set of steering, throttle, and braking commands, a humanoid with dexterous hands possesses up to 60 independent degrees of freedom that must be continuously balanced and synchronized in real time. Sim-to-Real Workflows and Unified World Models (14:22) Modern dynamic behaviors rely on capturing human motion and training reinforcement learning policies inside accelerated physics engines such as Isaac Gym and Isaac Lab before zero-shot deployment onto physical hardware. Although academic institutions have demonstrated impressive isolated skills—such as UIUC's autonomous fall-recovery policies and UC Berkeley's martial arts imitation—current learning pipelines remain sample-inefficient and fail to transfer physical intuition between disparate tasks like dancing and combat. Overcoming these limitations will require unified world models that treat dexterous manipulation and legged navigation as identical sensorimotor problems. Unitree estimates production-ready foundation world models remain three to five years away, pointing to early systems like Google's Gemini Robotics—which maps multi-modal video inputs directly into low-level joint trajectories—as the leading indicator of this architectural transition. While vertical manufacturing and physics simulation have driven humanoid hardware costs down to consumer-accessible levels, widespread utility remains blocked by fragmented control software. Until autonomous self-learning world models replace narrow reinforcement learning policies, humanoid deployment will remain confined to structured industrial workflows rather than general domestic environments.

精炼阅读摘要

宇树科技（Unitree Robotics）指出，人形机器人行业正从早期受限于高成本与机械瓶颈的实验室研发阶段，迈入由生成式人工智能驱动的具身智能竞速期。该机构认为，要真正实现物理通用人工智能（Physical AGI），必须突破离身智能缺乏物理实体的局限，依托自研垂直整合实现硬件降本，并最终走向统一的物理世界模型。 人形机器人演进阶段与场景落地路径 (0:04) 宇树科技将人形机器人的发展轨迹划分为三个时代。2021年之前为早期实验室阶段，以波士顿动力的Spot和本田的ASIMO为代表；受制于硬件瓶颈与高昂成本，仅有极少数机构具备制造能力，系统多作为昂贵的研究平台，缺乏商业可行性。2021年至2030年的第二阶段由生成式AI与基础模型突破引发，资本与产业投资大量涌入，通过注入高层推理能力推动具身智能研究竞赛。 宇树科技预测，2030年之后行业将迎来物理系统中的“真正AGI”。该机构预计其落地将分两步实现：首先部署于工厂制造与物流等结构化工业场景，随后逐步拓展至家庭等复杂的非结构化日常环境。 垂直整合降本与全尺寸人形硬件矩阵 (2:11) 宇树科技的硬件积累始于创始人王兴兴在2013至2016年硕士期间研发的XDog四足原型，并于2016年正式成立公司推进商业化。通过自主设计和制造高功率电驱动执行器、机械臂和定制激光雷达，宇树实施垂直整合供应链策略，在2020年将四足机器人成本降至1万美元以下，并相继推出消费级Go2以及工业巡检与救援导向的B2和轮腿版（Go2-W、B2-W）。 在人形机器人领域，宇树在2022年底启动项目后，两名核心工程师用约六个月时间便完成了全尺寸电驱人形机器人H1的原型开发，并在2024年初展示了电驱全尺寸人形机器人的后空翻动作及开源遥操作框架。随后推出的紧凑型G1高约1.3米、重35公斤、峰值关节扭矩达120牛·米，在现代物理仿真引擎优化下实现了1.4米的动态跳跃；依托成熟量产产线，宇树进一步推出了售价约6000美元的R1平台，以期降低学术研究门槛。 离身大模型的物理局限与学术运动探索 (8:53) 近期多所学术机构利用宇树硬件展开了密集验证：UIUC开发了全向自主跌倒起立策略，加州大学伯克利分校与清华大学训练了李小龙武术动作模仿，加州大学圣迭戈分校展示了舞蹈与灵巧手遥操作，斯坦福大学与伯克利实现了从单目人类视频中学习站立行走（VideoMimic），英伟达与卡耐基梅隆大学亦探索了多智能体篮球协作。 然而，宇树科技强调离身智能在物理世界中存在根本局限。纯语言模型控制物理身体就像“在梦境中摔倒，缺乏真实的身体感知”，无法仅靠神经元数量获得物理行动力。例如章鱼拥有超过5亿神经元，鹦鹉拥有2亿至30亿神经元，人类拥有160亿大脑皮层神经元，但鹦鹉因缺少手部无法洗衣做饭，猴子因发声器官限制无法发出人类语音。宇树认为，缺乏在真实物理实体上的直接交互与强化学习，机器人无法在现实环境中达到100%的任务成功率。 高维协同控制挑战与统一世界模型 (14:22) 人形机器人全身及灵巧手可达60个独立自由度，其实时协同控制的复杂度远超仅需映射转向与加减速的自动驾驶系统。宇树指出，当前AI基础模型在物理环境中面临三大结构性缺陷：一是缺乏自主自监督探索机制，高度依赖固化数据集；二是样本效率低且技能无法跨领域迁移，例如掌握舞蹈无法为学习武术提供物理直觉；三是操控与移动被割裂建模，而实际上二者均属于末端执行器与环境的物理交互。 破局之道在于构建能够像儿童通过试错学习抓取水杯那样、持续积累传感器运动经验的“世界模型”。在现有训练流程中，团队通常将动捕数据输入Isaac Gym或Isaac Lab等物理引擎进行强化学习，再零样本迁移至物理硬件。宇树评估认为，成熟实用的世界模型在量产机器人中的落地预计仍需3至5年时间；而谷歌Gemini等直接从多模态视频端到端输出底层关节控制轨迹的研究，已初步展现出下一代架构的发展方向。 从专用机械平台迈向物理通用人工智能仍面临样本效率低下、技能无法泛化以及高维动作协同困难等底层瓶颈。宇树科技与学术界的探索表明，软硬件垂直整合与大规模物理仿真大幅缩短了迭代周期，但具备连续自监督学习能力的实用化世界模型依然需要3至5年的技术演进与实体验证。

Detailed notes — zh

## Background 本视频由行业平台 Ai4 于 2025 年 8 月 29 日发布在 YouTube 上。内容为宇树科技（Unitree Robotics）北美区负责人 Tony Yang 的演讲，重点分享了人形与四足机器人的产业演进历程、宇树在硬件研发与规模化量产上的最新进展，以及构建物理世界大模型对实现具身智能与 AGI 的核心价值。 ## TL;DR 演讲系统梳理了人形机器人从早期探索、AI 驱动爆发到未来 AGI 落地工业与家庭的三阶段演进，并详细介绍了宇树从四足狗到全尺寸 H1、敏捷型 G1 及普及型 R1 的产品线突破。借助高保真仿真环境与 Sim-to-Real 强化学习，全球顶尖科研机构已在宇树硬件上实现了武术、篮球与自主爬起等高动态全身运动。Tony 强调大模型缺乏物理实体就像“在梦境中奔跑”，必须结合物理肉身与本体感知才能实现真实交互，并指出构建具备在线自学习与技能迁移能力的通用具身世界模型预计仍需 3 到 5 年。 ## Atmosphere & style 整场演讲节奏紧凑利落，兼具硬核工业量产的务实底色与前沿技术探索的远见视野。主讲人语言简练精悍，通过“在梦境中奔跑”和“生物神经元与身体结构”等生动具象的比喻，深入浅出地剖析了纯软件大模型与物理世界交互的本质脱节。现场氛围专业而充满张力，既直面了 Sim-to-Real 鸿沟与泛化迁移等行业痛点，又洋溢着对具身智能技术爆发式迭代的坚定信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) 人形机器人发展经历三个阶段：2021年前的早期探索、2021-2030年的AI驱动爆发，以及2030年后的真实AGI落地（先工业制造后家庭环境）。 - [02:11](https://www.youtube.com/watch?v=BMEsifCJL_c&t=131s) 宇树科技坚持核心零部件（执行器、机械臂、激光雷达等）自主研发与销售，并率先将四足机器人成本降至1万美元以内推向消费级市场。 - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) 随着生成式AI突破，宇树仅用2名工程师在6个月内研制出全尺寸电驱人形机器人H1，并在2024年实现全尺寸电驱空翻及开源遥操作系统。 - [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) 宇树相继推出高敏捷人形机器人G1（垂直弹跳达1.4米）与低至约6,000美元的轻量化人形机器人R1，依靠成熟供应链实现超低成本突破。 - [07:25](https://www.youtube.com/watch?v=BMEsifCJL_c&t=445s) 物理仿真环境（Genesis、Isaac Lab、MuJoCo等）的高保真建模极大缩短了运控软件迭代周期，Sim-to-Real强化学习成为主流开发范式。 - [07:25](https://www.youtube.com/watch?v=BMEsifCJL_c&t=445s) 全球顶尖高校基于宇树平台取得突破，如全姿态跌倒爬起算法，以及清华/UC Berkeley通过VideoMimic技术实现直接看人类视频学会行走与起立。 - [11:50](https://www.youtube.com/watch?v=BMEsifCJL_c&t=710s) 大模型与纯数字AI Agent缺少物理实体，如同“在梦境中奔跑”；具身智能必须直接在物理实体上运行强化学习以应对多关节毫秒级协调控制。 - [15:07](https://www.youtube.com/watch?v=BMEsifCJL_c&t=907s) 当前具身基座模型面临自学习泛化弱、Sim-to-Real虚实鸿沟、技能割裂无法迁移以及移动与操作模型割裂四大根本瓶颈。 - [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) 高难度动作编程流程主要依赖动捕提取人类轨迹、仿真器大规模并行强化学习训练，最后通过Sim-to-Real部署至物理实体。 - [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) 完全成熟的具身世界模型预计还需3到5年演进，端到端输入视频直接输出底层关节控制指令的模型展现出统一控制的巨大潜力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) 人形机器人行业的三阶段发展演进 宇树科技（Unitree Robotics）北美区负责人 Tony Yang 开启了关于人形与四足机器人行业现状与未来的分享。回顾人形机器人过去二三十年的历史，行业整体可以划分为三个核心发展阶段： 第一阶段为 2021 年之前的“早期探索阶段”。这一时期的代表包括波士顿动力（Boston Dynamics）的 Spot 狗以及本田（Honda）的 ASIMO 等先驱产品。虽然吸引了中美众多高校和企业参与研究，但全球能实现人形机器人量产的企业屈指可数，行业整体以实验室研发为主，硬件可靠性低且缺乏真实的商业落地场景。 第二阶段为 2021 年至 2030 年的“AI 驱动爆发阶段”。随着 ChatGPT 与生成式 AI（Generative AI）技术的爆发，大语言模型（LLM）与通用人工智能（AGI）理念被深度引入机器人领域。巨额资本涌入，产学研协同加速，机器人的智能化与运动控制水平迎来飞跃。 第三阶段为 2030 年之后的“真实 AGI 落地阶段”。届时人形机器人将分为两步全面融入现实世界：首先是进入工厂、制造等相对结构化的工业场景；随后迈入家庭等非结构化环境，实现真正意义上的通用具身智能。 ```mermaid flowchart LR accTitle: 人形机器人三阶段发展路线 subgraph S1["早期探索 (2021年前)"] A["ASIMO / Spot 探索"] --> B["高校实验室研究为主"] end subgraph S2["AI驱动爆发 (2021-2030年)"] C["ChatGPT/大模型注入"] --> D["巨额资本与量产尝试"] end subgraph S3["真实AGI落地 (2030年后)"] E["工业与制造场景应用"] --> F["家庭非结构化环境普及"] end B --> C D --> E ``` --- ### [02:11](https://www.youtube.com/watch?v=BMEsifCJL_c&t=131s) 宇树科技的产品布局与发展里程碑 宇树科技不仅聚焦四足机器人与人形机器人整机，还实现了核心零部件的完全自主研发与独立销售，包括自研执行器（Actuator）、机械臂及激光雷达（LiDAR）等，并延伸出工业巡检、消防救援等行业解决方案。回顾发展历程： - **2013-2016 年**：创始人读研期间设计并开发了初代原型 XDog。 - **2016 年**：创始人毕业正式创立宇树科技，推出第一代四足机器人。 - **2019 年**：发布第二代四足机器人 AlienGo，整机集成度大幅提升。 - **2020 年**：将四足机器人成本压缩至 1 万美元以内。 - **2021 年**：率先将四足机器人推向消费级市场，成为行业首个普及消费级四足狗的企业。 - **2023 年至今**：推出新一代消费级四足机器人 Go2 以及工业级重载四足机器人 B2，并拓展出轮足复合版 Go2-W 和 B2-W，广泛应用于电网巡检与特种消防场景。 --- ### [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) 人形机器人产品线迭代：H1、G1 与 R1 在 2022 年之前，由于生成式 AI 发展相对缓慢，宇树团队认为人形机器人并不适用于绝大多数商业落地场景。直到 2022 年底，随着 AI 技术取得关键突破，宇树果断立项人形机器人项目。 项目推进速度极具爆发力：仅由 2 名核心工程师在 6 个月内就研制出了全尺寸通用人形机器人 H1 的原型机。2023 年 3 月与 2024 年期间，团队高频迭代软件，并在 2024 年实现了全球首个全尺寸电驱人形机器人的原地空翻（Backflip），同年 8 月将遥操作控制软件全面开源。 ```mermaid flowchart TD accTitle: 宇树科技人形机器人迭代路线 subgraph H1["全尺寸旗舰 H1"] H1A["2023年初立项 / 6个月出原型"] --> H1B["2024年实现全电驱空翻"] H1B --> H1C["遥操作系统全面开源"] end subgraph G1["敏捷型 G1"] G1A["2024年5月发布 (高敏捷/轻量化)"] --> G1B["极限测试: 垂直起跳达1.4米"] G1B --> G1C["参数: 35kg / 1.3米高 / 峰值扭矩120N·m"] end subgraph R1["超低成本普及型 R1"] R1A["大规模量产经验迁移"] --> R1B["售价降至约 $6,000"] end H1C --> G1A G1C --> R1A ``` 2024 年 5 月，宇树发布了敏捷型人形机器人 G1。该机型高约 1.3 米，重约 35 公斤，最大关节峰值扭矩达 120 N·m，可选配三指灵巧手，并在 2024 年 10 月的软硬件极限测试中展现出高达 1.4 米的垂直弹跳能力。2025 年初，G1 在拳击、武术等复杂动态运动中持续突破。 此外，宇树近期推出了定价仅约 6,000 美元的轻量化人形机器人 R1。依托 G1 和四足狗成熟的规模化量产供应链，宇树将相同核心技术迁移至更轻量、少自由度的构型中，实现了前所未有的超低成本突破。 --- ### [07:25](https://www.youtube.com/watch?v=BMEsifCJL_c&t=445s) 仿真环境升级与学术界前沿突破 运动控制（Locomotion）软件迭代从过去按年计算缩短至一两个月一次，这得益于物理仿真器技术的飞速发展。过去模拟真实物理世界极其困难，而如今借助 Genesis、Isaac Lab、MuJoCo 等新型高保真仿真环境，摩擦力、重力等复杂物理接触因子能够得到高精度建模。宇树通过动捕（MoCap）采集人类动作数据，在仿真环境中训练强化学习策略（Policy），再通过 Sim-to-Real 技术无缝部署到实体机器人上。 全球顶尖高校与科研机构基于宇树 H1、G1 平台取得了一系列突破性进展： - **伊利诺伊大学厄巴纳-香槟分校（UIUC）**：研发出全姿态跌倒自主爬起算法，无论以何种姿态摔倒均能稳定复位。 - **加州大学伯克利分校（UC Berkeley）与清华大学**：让机器人学习李小龙武术动作；开发 **VideoMimic** 技术，突破了传统依赖昂贵遥操作采集数据的瓶颈，直接让机器人通过“看人类日常视频”学会行走、起立和坐下。 - **加州大学圣地亚哥分校（UCSD）**：利用灵巧手实现先遥操作后自主作业的泛化操作，并提出 **X-Robot/X-Body 2** 框架让机器人学习人类舞蹈与复杂全身动作。 - **斯坦福大学（Stanford）**：实现高精度人体动作模仿与全身动力学迁移。 - **英伟达（NVIDIA）与卡内基梅隆大学（CMU）**：联合训练机器人完成运球、投篮等高动态篮球技能。 --- ### [11:50](https://www.youtube.com/watch?v=BMEsifCJL_c&t=710s) 具身物理实体：大模型突破“梦境”的关键 大语言模型和纯数字 AI Agent 就像“在梦境中奔跑”——在梦里摔倒时，无论意识多清晰都会感觉“跑不快、使不上劲”，根本原因在于它们缺少一个物理实体（Physical Body）。 Tony 引用了生物学神经元数量对比实验：章鱼拥有超过 5 亿个神经元，非洲灰鹦鹉约 2 亿至 30 亿个，而人类大脑拥有约 160 亿个大脑皮层神经元。鹦鹉和猴子即便具备一定认知，受限于身体结构（如没有灵巧手或发声器官结构限制），永远无法完成洗衣服、做饭或人类语言交流。同样，人类即使智力再高，仅仅通过看视频也学不会骑自行车或开车，必须通过肉身在现实物理世界中摸索与试错。 具身智能的核心在于直接在物理实体上运行强化学习，实现复杂系统中 100% 的操作成功率。与自动驾驶汽车相对简单的输入输出不同，人形机器人拥有多达 60 个联动关节，需要在本体感觉（Proprioceptive Control）与外部多模态感知融合下实现高自由度的毫秒级协调控制。 --- ### [15:07](https://www.youtube.com/watch?v=BMEsifCJL_c&t=907s) 通用具身大模型（World Model）面临的挑战 要实现真正的具身 AGI，当前的基座模型与强化学习体系仍面临四大根本瓶颈： | 瓶颈维度 | 现状与核心问题 | 理想的具身 AGI 状态 | |:--- |:--- |:--- | | **自学习与泛化能力** | 严重依赖离线训练数据驱动，缺乏在线自学习能力，泛化性能脆弱。 | 能够自主理解与推理，利用已有先验知识自我迭代、越学越快。 | | **Sim-to-Real 鸿沟** | 强化学习训练效率低，虚拟仿真环境与真实物理接触之间存在显著 Sim-to-Real Gap。 | 高效跨越虚实鸿沟，在实体交互中实时闭环学习。 | | **经验与技能迁移** | 技能割裂（学跳舞和学功夫无法共享底层经验），无法像 LLM 那样积累并复用通用世界知识。 | 如同人类小孩学习抓杯子一样，无需重复编码即可将日常经验迁移至新技能。 | | **任务模型统一性** | 移动导航与灵巧操作使用割裂的模型。 | 统一大模型表达：本质上灵巧操作物体与底盘导航避障都是“机器人与外部环境的物理交互”。 | 构建一个具备实时在线感知与物理交互反馈的**世界大模型（World Foundation Model）**，是通向真实 AGI 的必经之路。 --- ### [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) 现场问答：动作控制流程与世界模型进展 在演讲结尾的互动环节中，Tony 针对现场观众关心的核心问题进行了详细解答： - **关于机器人高难度动作的底层编程流程**： 现阶段主要依赖动作捕捉系统（MoCap）精准提取人类专家动作（如功夫、舞蹈），将动捕轨迹数据输入 Isaac Gym / Isaac Lab 等高保真物理仿真器中进行强化学习策略（RL Policy）的大规模并行训练，完成域随机化收敛后，再通过 Sim-to-Real 部署至物理实体机器人上执行。 - **关于具身世界模型的研发进展与落地时间表**： 实现完全成熟的具身世界模型预计还需要 **3 到 5 年** 的演进时间。目前业内已出现类似的前沿探索，例如 Google DeepMind 推出的 Gemini / RT 系列具身模型，能够实现直接输入视频观察、端到端输出底层关节控制指令（Joint Control）来直接操控物体，展现出强大的统一端到端控制潜力。随着多模态大模型能力的飞速增强，具备物理常识与运动生成能力的世界模型将加速走向实用化。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) 第一阶段为 2021 年之前的“早期探索阶段”。 — Tony Yang > [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) 第二阶段为 2021 年至 2030 年的“AI 驱动爆发阶段”。 — Tony Yang > [00:04](https://www.youtube.com/watch?v=BMEsifCJL_c&t=4s) 第三阶段为 2030 年之后的“真实 AGI 落地阶段”。 — Tony Yang > [04:33](https://www.youtube.com/watch?v=BMEsifCJL_c&t=273s) 2024 年实现了全球首个全尺寸电驱人形机器人的原地空翻（Backflip），同年 8 月将遥操作控制软件全面开源。 — Tony Yang > [11:50](https://www.youtube.com/watch?v=BMEsifCJL_c&t=710s) 大语言模型和纯数字 AI Agent 就像“在梦境中奔跑”——在梦里摔倒时，无论意识多清晰都会感觉“跑不快、使不上劲”，根本原因在于它们缺少一个物理实体（Physical Body）。 — Tony Yang > [11:50](https://www.youtube.com/watch?v=BMEsifCJL_c&t=710s) 人类即使智力再高，仅仅通过看视频也学不会骑自行车或开车，必须通过肉身在现实物理世界中摸索与试错。 — Tony Yang > [11:50](https://www.youtube.com/watch?v=BMEsifCJL_c&t=710s) 具身智能的核心在于直接在物理实体上运行强化学习，实现复杂系统中 100% 的操作成功率。 — Tony Yang > [15:07](https://www.youtube.com/watch?v=BMEsifCJL_c&t=907s) 构建一个具备实时在线感知与物理交互反馈的世界大模型（World Foundation Model），是通向真实 AGI 的必经之路。 — Tony Yang > [17:34](https://www.youtube.com/watch?v=BMEsifCJL_c&t=1054s) 实现完全成熟的具身世界模型预计还需要 3 到 5 年 的演进时间。 — Tony Yang ## Mentioned Tony Yang, 宇树科技 (Unitree Robotics), 波士顿动力 (Boston Dynamics), Spot, 本田 (Honda), ASIMO, ChatGPT, XDog, AlienGo, Go2, B2, Go2-W, B2-W, H1, G1, R1, Genesis, Isaac Lab, Isaac Gym, MuJoCo, 伊利诺伊大学厄巴纳-香槟分校 (UIUC), 加州大学伯克利分校 (UC Berkeley), 清华大学, VideoMimic, 加州大学圣地亚哥分校 (UCSD), X-Robot/X-Body 2, 斯坦福大学 (Stanford), 英伟达 (NVIDIA), 卡内基梅隆大学 (CMU), Google DeepMind, Gemini / RT 系列 ## Video description Tony Yang, North American Director, Unitree Robotics Abstract With artificial intelligence technology move very fast in the past two years, humanoid robot have been one of the most import form to realized embodied AI and AGI, Unitree have been working for more than 8 years in leg robot and 1 year in humanoid robot area. There are three most important parts, algorithm, data and computing capability for realized AGI. Those three part will finally running on physical robots, we believe build robust physical humanoid robot system is key for this ecosystem, and World Large-Scale Model (most people called foundation model) is the key to bring Embodied AI for for humanoid robot, we will share the most important progressing have been made on industry and research side in the past one year, and expect and excited for new progressing will happening in next few years soon. [Source](https://www.youtube.com/watch?v=BMEsifCJL_c)

Source quotes

[0:04] Tony Yang: 第一阶段为 2021 年之前的“早期探索阶段”。 [0:04] Tony Yang: 第二阶段为 2021 年至 2030 年的“AI 驱动爆发阶段”。 [0:04] Tony Yang: 第三阶段为 2030 年之后的“真实 AGI 落地阶段”。 [4:33] Tony Yang: 2024 年实现了全球首个全尺寸电驱人形机器人的原地空翻（Backflip），同年 8 月将遥操作控制软件全面开源。 [11:50] Tony Yang: 大语言模型和纯数字 AI Agent 就像“在梦境中奔跑”——在梦里摔倒时，无论意识多清晰都会感觉“跑不快、使不上劲”，根本原因在于它们缺少一个物理实体（Physical Body）。 [11:50] Tony Yang: 人类即使智力再高，仅仅通过看视频也学不会骑自行车或开车，必须通过肉身在现实物理世界中摸索与试错。 [11:50] Tony Yang: 具身智能的核心在于直接在物理实体上运行强化学习，实现复杂系统中 100% 的操作成功率。 [15:07] Tony Yang: 构建一个具备实时在线感知与物理交互反馈的世界大模型（World Foundation Model），是通向真实 AGI 的必经之路。 [17:34] Tony Yang: 实现完全成熟的具身世界模型预计还需要 3 到 5 年 的演进时间。

Source quotes

[0:00] Tony Yang: Unitree have been working for more than 8 years in leg robot and 1 year in humanoid robot area. [0:00] Tony Yang: There are three most important parts, algorithm, data and computing capability for realized AGI. [0:00] Tony Yang: build robust physical humanoid robot system is key for this ecosystem, and World Large-Scale Model (most people called foundation model) is the key to bring Embodied AI for for humanoid robot [8:53] Tony Yang: falling down in a dream [8:53] Tony Yang: you do not have a physical body [8:53] Tony Yang: you always cannot feel that you can run fast. [8:53] Tony Yang: VideoMimic [14:22] Tony Yang: World Models

Original transcript

[0:04] Hello everyone. Uh my name is Tony. I'm [0:06] from Uni Robotics responsible for North [0:08] American business. Uh today I'm going uh [0:12] give some insights about human robot and [0:14] the quant robot industry. So for the [0:17] human robot actually we in the past 20 [0:20] to 30 years we already have a long time [0:23] history. At the moment we think the [0:25] human robot probably will be have a [0:27] three stage. So the first stage is we [0:30] called earlier period which is before [0:32] 2021 at that stage. So like a uh spot [0:36] and from Boston dynamics and uh even the [0:39] ashimo from like a hundai. So those [0:43] robots already uh coming and uh I push a [0:47] lot of company and university start [0:49] doing join the research but even in [0:52] China and the US. So it's still only few [0:54] company which can do the massive [0:56] production for for the human robots and [0:59] this parade actually mostly for research [1:02] rather than real application scenario [1:04] still have a lot of hardware program and [1:06] after 2021 so uh one thing happening is [1:10] that the chart GBT is coming and the [1:12] generative AI help uh push the robotics [1:16] industry a lot. So we think from 2021 to [1:20] 2030 [1:21] approximately 10 years. So the robot the [1:25] the human robot will be like a [1:28] coming robotics industry get a huge [1:30] investment which is happening now and [1:33] the more university and the company will [1:35] be getting involved and also the more [1:38] intelligence on the robots with large [1:40] language model and also AGI conception. [1:43] after 2030 we think uh the real AI will [1:47] be coming. So we'll be still have a two [1:49] stage. So the first stage will be most [1:52] of robot probably will be used uh uh [1:55] scenario which is like a factory or like [1:58] a manufacturing and an manufacturing [2:01] industry and the second stage will be [2:03] like a unstructed scenario like even for [2:06] home use purpose. So that that will be [2:08] the real AGI which is coming [2:11] and for unry we actually uh this is all [2:14] the product which we have. So we do not [2:17] only uh focus on the quantity robot but [2:20] also focus on humanoid robot uh all the [2:24] robots actuator which we soft design and [2:28] soft uh dur during the soft research and [2:30] development and we create the not only [2:34] those pro uh those robots but we also [2:37] saw in the activator separately as well [2:40] and based on the quant robot and human [2:43] robot. So we did some solution like the [2:45] photo quantity robot inspection and fire [2:47] rescue solution. We even use activator [2:50] technology to create the robot arm and [2:52] also like a lighter those different [2:55] product. But back to uh more than 10 [3:00] years ago. So in 2013 to 2017 16 when [3:04] our founder he studying university for [3:06] his master degree. So his design and [3:08] developed the XTO and then on the 2016 [3:12] so after he graduated from university so [3:15] he established the uni robotics and then [3:17] start build the first generation robot [3:19] and then uh to the 2019 so when we [3:23] release the the second generation robot [3:25] alien go which is uh makes the robot [3:28] much more compatible than the older [3:30] generation and then to the 2020 so which [3:33] will reduce the cost for the quantum [3:36] robot less than $10,000 and then 2021. [3:40] So we bring the quant robot to the [3:43] consumer level which is the first [3:45] company during the consumer level quant [3:47] robot and also then to the 20 20 uh 2030 [3:52] 2023. So which is we started the new [3:54] generation robot which is the goto robot [3:59] and and that's still the uh latest robot [4:02] which we have. [4:09] So and also for the industry robots so [4:13] which we did it's on the B2 robots and [4:16] also for the uh world leg robots on the [4:18] goto as well and then we we did the well [4:22] robot on the B2 for the industry as well [4:24] and then for the uh fire rescue and also [4:27] for the inspection solution as well. [4:33] So for human robot which we did we [4:36] started the project which uh at the [4:38] earlier of the 2023 the reason we during [4:43] the human robot is because before 2022 [4:47] when generative AI it's move very slowly [4:50] before 2022 so we think human robot it's [4:54] for not most of scenario it's not [4:56] suitable for human robot and then after [4:58] 2020 at the end of the 2022 so we made [5:02] the decision to start the human robot [5:04] project and in approximately six months. [5:06] So we we have a two engineer and in six [5:10] months build the prototype of H1 cumin [5:12] robot and then in the uh 20 in the 2023 [5:18] on the on the uh on the March so which [5:21] we upgrade the software and then on the [5:24] 2024 so uh we upgrade the software again [5:28] we even did the on 2024 we did the first [5:31] backflip 40 fulls size electronic [5:34] solution human robot [5:37] And then uh last year August so we also [5:40] did remove operation software program [5:43] for full for the open source for [5:45] customer [5:49] and the G1 robot which we start from [5:52] last year in the May release the first [5:56] uh robots small size but this robot is [5:59] very agile which you already saw [6:01] probably saw on the online and also uh [6:05] like on the uh exhibition hall probably [6:08] on the yesterday and this one. So in [6:12] last year October we upgrade the [6:14] software again which increase the [6:18] stability of the locomotion and also [6:20] test some limits of this hardware. So [6:23] the hardware shows very stronger for a [6:26] lot of uh like movements which is [6:30] already over expected like the jump. [6:33] So the jump is going up to 1.4 four [6:36] meters and even some very stronger [6:39] movements but the hardware still can [6:41] handle those type of movements [6:45] and all the movements which we collect [6:47] the data from human and then uh we use [6:50] those data put in the simulation train [6:52] the policy and then did the same to real [6:55] to the physical robots [6:57] and this is the this year on the January [7:00] so we did the upgrade software uh [7:03] upgrade again and even from February to [7:06] the to the March. So if you look at all [7:10] the upgrade which almost in one or two [7:13] months we make the software especially [7:17] for the locom motion software upgrade [7:20] may make make a very uh move very fast [7:23] on the locomotion side. The reason is [7:25] because uh we already saw the generative [7:27] AI help the locom motion of the robots [7:32] move very fast especially in the [7:33] simulation environment. So formally we [7:35] we think the like a to simulate a [7:38] physical world is very hot. Now it's [7:41] like a genesis or is lab or like a [7:44] podit. So with those type of simulator [7:48] it's really can simulate the like a [7:51] fraction or like the gravity also those [7:55] physical uh sec f factor very well and [8:00] on the April so this year we did a [8:03] boxing I believe uh you probably saw the [8:07] boxing robot already and the robot [8:09] actually also have a three fingers dex [8:11] hand as optional with 35 kilogram and [8:14] the maximum head is about 1.3 mters. Uh [8:18] maximum joint talk is up to 120 nomes. [8:22] And this is a new model which we [8:24] released uh like a few days ago. [8:28] Uh the robot is only cost about $6,000. [8:32] So the reason we can drop down the cost [8:35] a lot is because we already on the [8:37] massive production stage and uh we we [8:41] are just shift uh same technology to the [8:45] smaller robots. So the R1 robot is a [8:47] little bit smaller than the G1 criminal [8:50] robot and less degra [8:53] and some research which we already [8:55] noticed from like a different [8:57] university. So they did a lot of [8:59] research uh on the different robots [9:02] especially like on the locomotion sides [9:05] also on the like uh like this one. So [9:08] this is come from the uh university of [9:11] Illinois uber champion. So they did a [9:14] research which is no matter what what [9:16] what uh what type of posture when the [9:19] robot is fall down so the robot can [9:21] still uh stand up from the ground and [9:25] this is like the robot learning the [9:29] Bruce Lee like for the [9:32] uh different for different movements [9:35] which is from the UC Berkeley and also [9:37] the Tinuan University and this is from [9:40] UCSD. So the with the next hand so the [9:44] robot able to manipulate the different [9:46] task. So firstly they will teleoperate [9:48] the robot and then go to the autonomous [9:51] mode. So the robot able to autonomously [9:54] during some of task and some of the like [9:58] this is from Berkeley. So the use [10:01] integrated with language model to the [10:03] human robot as well and the Stanford. So [10:06] which is also use the H1 human robot [10:09] which im imitate the human uh movements [10:13] and shifted the human movements to the [10:15] uh human robot as well. And this is also [10:18] kind of the research from universal [10:21] Berkeley. So they called the uh video [10:25] mimic. So the video mimic is a uh [10:28] firstly you you fed the robots a video. [10:31] The biggest challenge for the human [10:33] robot at the moment it's it's not like a [10:35] large language model. So we have a huge [10:38] text video resource data from the [10:41] network we can acquire. But the for [10:44] human robot normally we have to [10:47] teleoperate uh teleopterate the robot to [10:49] collect data. Now they can feed the only [10:52] the video to the robot and the robot can [10:54] learn the like the working scale or like [10:57] a stand up or sit down those type of [11:00] skill only from video. So this is a [11:03] really interesting research I mean that [11:06] that is happening recently and some uh [11:09] like this is from the invadia and also [11:11] uh semu which they did like a play b [11:15] basketball basketball and then to train [11:18] this robot have those type of skill and [11:21] this is also from Berkeley which is [11:23] doing a lot of like sim simulator to [11:26] improve the performance of the simulator [11:30] and also for the X body 2 which is from [11:33] the USSD. So learning uh teach the robot [11:36] to do the dance and also different [11:38] movements. So if you look at all the [11:41] latest research and all the advanced [11:44] research you already saw some like [11:48] uh [11:50] those a lot of a lot of task is only [11:53] happening in the lab. So uh when I was [11:56] young I also um like when fall for fall [12:01] fall down in the dream so I feel I [12:03] cannot run very fast. So the reason is [12:05] because the currently humanoid robot or [12:08] other other robots it's because like the [12:11] the fall down in like the large language [12:14] model is like a fall down in the dream [12:16] because they do not have a physical [12:18] body. So if on the dream so you do not [12:20] have physical body you always cannot [12:22] feel that you can run fast. [12:26] And that's the that's the that's the [12:28] problem at the for the large language [12:30] model or any AI agent. So because they [12:33] do not have a language uh physical body. [12:36] So we think physical body is very [12:37] important for the for the AI to do [12:40] physical task. Uh one of the interesting [12:44] experiment we are thinking about if we [12:46] can change the brain between like a pyro [12:49] or human or like a monkey uh between [12:52] monkey and human. So like for the [12:55] neurosur octopus so that they have more [12:58] than 500 million euros like a paras [13:01] approximately 200 millions to three [13:03] billions but human up to 16 billion [13:08] uh for human to do a type different task [13:11] like to do laundry or like a cooking so [13:15] which is very easy but for parro for [13:17] monkey it's it's almost impossible [13:20] because like parro they do not have a [13:22] hands so it's very hard to do manipulate [13:25] this this type of task and for monkey so [13:28] they can learning skill uh to speak but [13:31] it's still very hard to like uh because [13:35] they have the structure to to to to [13:37] speak that's the reason so even with [13:40] human intelligence we cannot learning [13:42] how to ride a bicycle not only through [13:44] the watching video still need practice a [13:47] lot that's that is like a when you do [13:50] the uh we do the reinforcement learning [13:52] So we have to teach the robot on the [13:55] physical robots to make the robot learn [13:57] those skill. [13:59] It's like a human. So you cannot just [14:02] watch the video to learn a skill like a [14:04] drive car and the importance of directly [14:07] reinforcement learning for physical [14:08] robots for especially for the complex [14:11] robot system to acquire 100% successful [14:14] rate. So the currently problem for the [14:16] cumulative robot is still the successful [14:19] rate for different task [14:22] and also for the control side. So pro [14:25] perceptive control or like a control [14:27] based on the sensor which is still like [14:29] a which which one is better. We still [14:32] have a like exploring this questions. [14:36] uh and this dex hand which we have is a [14:39] five fingers Dex hand which is we add [14:41] mod but the biggest challenge for the [14:44] cumul robot which we have is like a [14:46] compared with autonomous vehicle so it's [14:49] very different autonomous vehicle [14:50] normally you have a very simply import [14:52] and output but for human robot you have [14:56] up to 60 joins which means you can you [14:59] have to control the 60 join at the same [15:01] time so it's much more than complicated [15:04] than autonomous vehicle [15:07] So some challenges of the real AGI we [15:09] think the AI foundation model is do not [15:12] have a self-arning capability at the [15:14] moment. So all the foundation model [15:15] structure is driven by trained data. The [15:18] capability re very relied on the data [15:21] and made it very weak for gen [15:23] generalized and the currently uh like a [15:27] reinforcement learning basically because [15:28] the efficiency is still very weak and [15:31] you you need to do huge training on the [15:33] simulation when during semi real so they [15:36] have a big gap between the sim [15:37] simulation and to the physical robots [15:40] and each movements which we train at the [15:42] moment we found so you train the robot [15:45] to do dance you train the robot to do [15:47] kung fu you have to you you cannot do [15:50] the like a uh use the existing [15:53] experience or knowledge to teach the [15:56] robots. So the robot cannot acquire the [15:58] experience or the knowledge. It's not [16:00] like a language model. And third is the [16:03] model we use at the moment is not an [16:06] unification model for example. So the [16:08] dexand manipulation essentially is [16:11] similar as a navigation. So when the [16:13] robot is during navigation, it's [16:15] interact with the environment, but it's [16:18] essentially the same as a dex hand when [16:20] the do the manipulation is also interact [16:22] with the the object as well. And the [16:25] perfect model for AI we think is still [16:28] continue learning and by himself [16:31] understanding the reasoning use the [16:33] existing knowledge or learning new [16:35] knowledge get faster and faster [16:36] learning. This is how the human thinking [16:39] and the learning because if you learn [16:41] some skill so you still can have [16:44] existing occupied the experience. So a [16:47] lot of skill which we have it's also [16:50] very relied on experience. Imagine when [16:53] is the kids to pick up the button or or [16:57] pick up a cup. So there's no nobody [17:00] probably teach him how to eat how to [17:02] pick up the cup but it's rely on the [17:05] experience. [17:08] And we think the probably word model [17:11] will be the next like to do to to for [17:15] the real AGI. So the uh but it requires [17:19] physical robots to physical deploy the [17:21] real time to acquisition up to date [17:24] training also need to engage the [17:26] physical interaction in the human life. [17:31] Yeah. Uh any question? [17:34] >> Thank you Tiffany. We have I think about [17:37] one or two questions that Tony can [17:39] address. Anyone? [17:44] This is your last chance. [17:47] >> Two minutes. [17:50] >> Uh yes. So yes. So to normally program [17:54] the movements of the robots at the [17:56] moment uh firstly we have a motion [17:58] capture system especially to do the gong [18:01] for dance. So you have to capture the [18:02] movements from human and then use those [18:04] data put in the simulator like as lab as [18:08] gym to do the to train your policy and [18:11] then after train the policy you can [18:13] deploy the policy from sim uh from to do [18:16] from sim to real to the physical robots. [18:22] >> Okay we have one more question right [18:24] here. [18:27] >> Okay. So um you mentioned at the end [18:30] about the world model so like uh what's [18:33] the progress on it like how far are we? [18:36] Yeah, very good question. I I think they [18:38] still need approximately three to five [18:40] years. So they already have some very [18:43] similar world model is it's coming like [18:46] a uh gini jini robotics model from [18:50] Google which is really just feed the [18:52] robot like some video and the directly [18:55] output the join join the control so [18:58] control the joints directly to [19:00] manipulate the the object which is very [19:03] impressive [19:05] and the generative of this Germany [19:07] robotics model like a Germany uh three. [19:10] So I think it's enhanced a lot.