Unitree Ships 11,000 Humanoids as Low Prices Drive Global Sales

Building the Next-Generation Humanoid Robot and Ecosystem with Unitree Robotics — Ai4

Ai4

Unitree Robotics is betting an open software ecosystem and sub-$5,000 hardware will secure market dominance over specs-focused humanoid competitors.

AI-assisted summary of Building the Next-Generation Humanoid Robot and Ecosystem with Unitree Robotics — original source. Check the original for full context.

Source published:

Detailed notes

## Background This video was published on YouTube by the channel Ai4 on August 21, 2026. The description notes that Nick Xu, Key Account Lead at Unitree Robotics, introduces Unitree's product offerings and future roadmap.

## TL;DR Nick Xu outlines Unitree Robotics' transformation from a quadruped manufacturer into a leader in humanoid robotics, highlighting models like the H1, the bestselling G1, the heavy-duty H2, and the ultra-affordable $4,900 R1. He details how simulation tools, teleoperation, and open-source models like 'Uni-World' enable autonomous real-world operations and in-factory self-assembly. Furthermore, Xu explains how cultural milestones, high-flying acrobatic achievements, and a strong global research network contribute to an ecosystem-driven competitive moat.

## Atmosphere & style The presentation carries a professional, forward-looking, and ambitious tone suitable for an industry keynote. Nick Xu presents a technical yet accessible narrative, balancing commercial milestones and hardware specs with impressive feats of acrobatic engineering and real-world deployment. Viewing the session feels like gaining an insider's look into the rapid evolution and ecosystem expansion of cutting-edge robotics.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) Unitree expanded from quadruped robotics into humanoid platforms due to rising customer demand for affordable research systems and breakthroughs in generative AI. - [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) Unitree's humanoid portfolio includes full-size bipeds (H1, H2, H2+), compact bipeds (G1, R1), and hybrid wheeled platforms (G1D, R1D, H2D) mounted on AGV bases. - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) The G1 humanoid is Unitree's bestselling model, exceeding 5,500 units sold in 2025 and over 11,000 total units produced. - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) Unitree uses simulation software such as NVIDIA Isaac Sim, MuJoCo, and Gazebo to train locomotion policies via reinforcement learning before deploying them to real hardware. - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) The wheeled G1D platform functions as an embodied AI data collection system that uploads training data to private clouds or public infrastructure like AWS. - [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) The full-size H2 platform features 31 degrees of freedom, weighs roughly 150 lbs, and offers onboard compute options including Intel Core i7 and NVIDIA Jetson AGX Orin. - [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) Unitree collaborated with NVIDIA to build the H2+ reference platform, which features Schunk dexterous hands for high-level development. - [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) Unitree offers proprietary end-effectors (Dex5, Dex3-1, Dex1) and supports third-party dexterous hands from manufacturers like Ouji, Robotiq, Inspire-Robots, and Schunk. - [07:26](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=446s) The ultra-affordable R1 humanoid starts at around $4,900 and was selected as one of TIME Magazine's Best Inventions of 2025. - [07:26](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=446s) Unitree aims to automate its own factory production lines using humanoid robots trained via teleoperation to assemble electric motors. - [09:46](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=586s) Cultural appearances at the CCTV Spring Festival Gala generated massive exposure and spurred the founding of over 100 humanoid robotics startups in China. - [09:46](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=586s) Reinforcement learning advances enabled Unitree robots to achieve world records, including continuous freestyle table tennis rallies and 3-meter-high aerial flips. - [09:46](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=586s) Unitree's open-source Uni-World model bridges the sim-to-real gap to automate real-world tasks across household, healthcare, warehouse, and agricultural settings. - [13:36](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=816s) More than 200 published academic research papers utilize Unitree hardware, alongside partnerships with major tech firms like NVIDIA and Scale AI. - [13:36](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=816s) Unitree's primary long-term competitive moat is an open, interconnected hardware and developer ecosystem similar to Apple and NVIDIA.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) From Quadruped Pioneers to Humanoid Robot Leaders

Nick Xu, Key Account Lead at Unitree Robotics, opens his presentation by detailing the company's evolution from a quadruped (robot dog) manufacturer into a full-fledged humanoid robotics pioneer. Initially, Unitree established its market presence by shipping quadrupedal robots across educational, industrial, and consumer sectors globally. However, the surge in generative AI, catalyzed by ChatGPT, combined with strong customer demand for reliable and reasonably priced research platforms, propelled Unitree into the humanoid space.

They introduced their first full-size humanoid, the H1, in 2023. This was followed in 2024 by the compact G1 humanoid, and in 2025 by the highly affordable R1. To serve applications where wheeled stability outweighs bipedal agility, Unitree developed hybrid wheelbase models (such as the G1D, R1D, and H2D) that mount the upper torso of their bipedal robots onto Automated Guided Vehicle (AGV) bases. Their full-sized line also evolved with the release of the upgraded H2 and H2+ models.

- **Foundational Lineup:** Started in quadruped robotics serving consumer, education, and industrial sectors. - **H1 Humanoid:** Unitree's initial full-sized bipedal humanoid, launched in 2023. - **Product Expansion:** Expanded into smaller bipeds (G1, R1), full-sized upgrades (H2, H2+), and AGV-mounted wheeled hybrids (G1D, R1D, H2D).

---

[02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) The G1 Best-Seller and Wheeled Data-Collection Platforms

Unitree's G1 humanoid—originally unveiled in May 2024 at the IEEE International Conference on Robotics and Automation (ICRA)—has become their most iconic and bestselling model. Designed initially as a pure research platform, the G1 achieved massive popularity: Unitree sold over 5,500 units globally in 2025 alone, pushing total cumulative production past 11,000 units. To power its dynamic mobility, locomotion policies are trained using reinforcement learning in simulation software like NVIDIA Isaac Sim, MuJoCo, and Gazebo before being deployed onto real hardware.

Equipped with 3D LiDAR and depth cameras on its head, the G1 supports compute modules ranging from the NVIDIA Jetson Orin NX to back-mounted AGX Orin units. Building on this platform, Unitree launched the wheeled G1D to serve as a high-efficiency data collection platform for embodied AI, enabling data uploads to private clouds or public infrastructure like AWS. The wheeled design offers enhanced stability, making it ideal for environments like warehouses, assembly lines, energy plants, beverage service, and supermarkets.

- **Sales Volume:** Over 5,500 G1 units sold in 2025; cumulative production exceeded 11,000 units. - **Simulation Tools:** Locomotion policies trained in NVIDIA Isaac Sim, MuJoCo, and Gazebo. - **G1 Compute Options:** Head-mounted 3D LiDAR and depth camera; compute options include NVIDIA Jetson Orin NX and AGX Orin. - **Embodied Data Collection:** Wheeled G1D platform collects training data for deployment across AWS or local cloud environments.

---

[05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) Full-Size Platforms, Hardware Partnerships, and Dexterous Hands

For heavy-duty applications, Unitree offers the full-scale H2 humanoid platform, featuring 31 degrees of freedom (DoF) and weighing roughly 150 lbs. Powered by onboard compute options including Intel Core i7 processors and NVIDIA Jetson AGX Orin modules, the H2 is targeted at industrial automation, commercial service, and high-level research. Unitree also partnered with NVIDIA to develop the H2+, a reference platform equipped with Schunk dexterous hands designed specifically for advanced robotics developers.

Alongside their bipedal and wheeled form factors, Unitree has expanded its end-effector ecosystem. They offer proprietary end-effectors like the Dex5 hands (compatible with the H2), the Dex3-1 hands (compatible across R1, G1, and H2), and Dex1 grippers. To provide maximum flexibility for researcher needs, Unitree platforms seamlessly integrate with third-party dexterous hands from manufacturers like Ouji, Robotiq, Inspire-Robots, and Schunk.

- **H2 Specifications:** 31 Degrees of Freedom (DoF), ~150 lbs, with Intel i7 and NVIDIA Jetson AGX Orin options. - **NVIDIA Collaboration:** H2+ platform developed alongside NVIDIA using Schunk dexterous hands. - **Proprietary End-Effectors:** Dex5, Dex3-1, and Dex1 grippers. - **Third-Party Hand Integration:** Compatible with hands from Ouji, Robotiq, Inspire-Robots, and Schunk.

---

[07:26](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=446s) Ultra-Affordable R1 and In-House Factory Automation

To democratize access to humanoid robotics, Unitree introduced the small-scale R1, which Nick calls "the most cost-effective humanoid robot" on the market. Starting at approximately $4,900, the Iron Man-inspired design earned recognition on *TIME* Magazine's list of the Best Inventions of 2025. Like the rest of the lineup, the R1 torso can also be mounted on an AGV base (R1D) for wheeled deployment.

A core long-term objective for Unitree is achieving full self-manufacturing automation—using humanoid robots to assemble other humanoid robots. In-factory footage demonstrates G1 units executing precise sorting tasks and assembling motor modules for their own production lines. These complex manipulation skills are trained using teleoperation, where human operators demonstrate assembly procedures that the AI models continuously learn from.

- **R1 Pricing:** Starting around $4,900; recognized as one of *TIME* Magazine's Best Inventions of 2025. - **Self-Manufacturing Vision:** Robots training to assemble Unitree's own electric motors on active production lines. - **Training Method:** Human operator teleoperation used to gather demonstration data for assembly tasks.

---

[09:46](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=586s) Cultural Milestones, High-Flying Acrobatic Records, and Open Source Models

Humanoid robotics achieved a landmark cultural breakthrough in China when six Unitree H1 robots performed a coordinated group dance during the 2024 CCTV Spring Festival Gala, generating millions of views. The performance ignited widespread investment across the nation, sparking the founding of over 100 new humanoid robotics startups within a single year. Unitree returned to the Gala in 2025 with their "WuBot" performance, showcasing dramatic breakthroughs achieved via reinforcement learning.

These advancements include breaking multiple world records, such as performing the world's first continuous freestyle table tennis rally by a humanoid and executing a 3-meter-high aerial flip alongside multiple continuous backflips. Behind these achievements is Unitree's open-source "Uni-World" world model. The model enables fully autonomous execution of real-world tasks trained in simulation—ranging from household chores like dishwashing and laundry to healthcare assistance, supermarket sorting, warehousing, and agricultural work.

- **Spring Festival Gala Impact:** 2024 performance by six H1 robots catalyzed the founding of 100+ humanoid startups in China. - **World Record Feats:** First continuous freestyle table tennis playing and 3-meter-high continuous aerial backflips. - **Sim-to-Real Progress:** Reinforcement learning continuously shrinks the gap between physics simulations and real-world execution. - **Uni-World Model:** Open-source world model demonstrating autonomous laundry, dishwashing, healthcare aid, and warehouse operations.

---

[13:36](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=816s) The Ecosystem as a Competitive Moat

Unitree's global footprint spans top technology companies like NVIDIA and Scale AI, alongside hundreds of academic institutions worldwide. Over 200 scientific research papers have already been published using Unitree hardware as a foundational platform. Facing intense domestic and global competition, Nick stresses that product specs alone are not enough to maintain leadership.

Drawing comparisons to tech giants like Apple and NVIDIA, Nick explains that building a sticky, open ecosystem is Unitree's primary competitive moat. By continuously releasing reliable, cost-effective hardware platforms and supporting an open-source developer ecosystem, Unitree ensures that researchers, developers, and enterprise clients remain embedded in their environment for years to come.

- **Academic Impact:** Over 200 published research papers based on Unitree humanoid platforms. - **Major Partners:** NVIDIA, Scale AI, and leading global research universities. - **Strategic Vision:** Establishing an ecosystem-driven competitive moat akin to Apple and NVIDIA.

## Notable quotes > [07:26](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=446s) the most cost-effective humanoid robot — Nick Xu

## Mentioned Nick Xu, Unitree Robotics, ChatGPT, H1, G1, R1, G1D, R1D, H2D, AGV, H2, H2+, IEEE International Conference on Robotics and Automation, ICRA, NVIDIA Isaac Sim, MuJoCo, Gazebo, NVIDIA Jetson Orin NX, NVIDIA Jetson AGX Orin, AWS, Intel Core i7, NVIDIA, Schunk, Dex5, Dex3-1, Dex1, Ouji, Robotiq, Inspire-Robots, TIME Magazine, Best Inventions of 2025, CCTV Spring Festival Gala, WuBot, Uni-World, Scale AI, Apple

## Video description Nick Xu - Key Account Lead, Unitree Robotics

I will introduce about Unitree's Product and future roadmap.

[Source](https://www.youtube.com/watch?v=7mZvwFtF3vU)

中文简报

低价驱动全球销量，宇树人形机器人出货达1.1万台 宇树科技押注开放软件生态与5,000美元以下的硬件，以在与侧重硬件参数的人形机器人竞品的竞争中确立市场主导地位。 Nick Xu 概述了宇树科技（Unitree Robotics）从四足机器人制造商转型为人形机器人领军企业的历程，重点介绍了 H1、畅销款 G1、重载款 H2 以及售价仅 4,900 美元的超平价 R1 等机型。他详细阐述了仿真工具、遥操作以及“Uni-World”等开源模型如何赋能现实场景中的自主作业与工厂内自我组装。此外，Xu 还解释了文化里程碑、高难度特技动作以及强大的全球研究网络如何共同构建起由生态系统驱动的竞争护城河。 Unitree 从四足机器人拓展至人形机器人领域，产品矩阵现已涵盖全尺寸双足（H1、H2、H2+）、小型双足（G1、R1）以及搭载 AGV 底盘的轮式混合平台（G1D、R1D、H2D）。 (0:04) 小型人形机器人 G1 成为 Unitree 最畅销的旗舰型号，2025 年全球销量突破 5,500 台，累计产量已超 11,000 台，其轮式版本 G1D 可作为具身智能数据采集系统将训练数据回传至云端。 (2:34) 全尺寸 H2 平台具备 31 个自由度，并联合 NVIDIA 打造了配备 Schunk 灵巧手的 H2+ 参考平台；同时 Unitree 还推出了自研灵巧手并全面兼容第三方末端执行器。 (5:56) Unitree 推出了起售价约 4,900 美元的极具性价比人形机器人 R1，入选《时代》周刊 2025 年最佳发明，并在工厂内探索使用 G1 机器人自组装电机等组件以实现自制造愿景。 (7:26) 基于强化学习与开源 Uni-World 世界模型，Unitree 人形机器人创下了连续乒乓球对打及高空翻转等纪录，并依托与 NVIDIA、Scale AI 等伙伴的合作构建类似苹果的开发者生态护城河。 (9:46)

Reading summary

Nick Xu, Key Account Lead at Unitree Robotics, outlines the company's transition from quadruped robotics to a broad humanoid lineup, arguing that sustainable market leadership in embodied AI relies on building an open, accessible developer ecosystem rather than competing on hardware specifications alone. Hardware Scaling and Hybrid Wheeled Formats (0:04) Following initial global sales of quadrupedal robots, Unitree expanded into humanoid robotics in response to surging generative AI interest and market demand for accessible research hardware. The company introduced its full-size H1 in 2023, the compact G1 in 2024, and the entry-level R1 in 2025 at approximately $4,900. By 2025, annual sales of the G1 exceeded 5,500 units, pushing cumulative production across the platform past 11,000 units. To address industrial scenarios where bipedal mobility is unnecessary, Unitree developed hybrid wheelbase variants (G1D, R1D, and H2D) that mount bipedal torsos onto Automated Guided Vehicle (AGV) bases. Equipped with 3D LiDAR, depth cameras, and onboard compute modules such as the NVIDIA Jetson Orin NX or AGX Orin, these wheeled units function as stable data-collection platforms for training embodied AI models across warehouses, supermarkets, and energy facilities. Dexterous Manipulation and In-Factory Automation (5:56) For heavier industrial and research deployments, Unitree developed the full-scale H2, a 150-pound humanoid with 31 degrees of freedom powered by Intel Core i7 processors and NVIDIA Jetson AGX Orin compute. The platform is complemented by the H2+, a developer reference model built in partnership with NVIDIA and fitted with Schunk dexterous hands. Across its portfolio, Unitree supports proprietary end-effectors—including the Dex1, Dex3-1, and Dex5—as well as third-party hands from manufacturers such as Ouji, Robotiq, and Inspire-Robots. A central operational goal for Unitree is self-manufacturing automation, using humanoid units to build their own hardware components. In factory trials, G1 robots perform component sorting and assemble electric motor modules on active production lines. These fine-motor skills are developed through teleoperation, where human operators demonstrate assembly routines to supply imitation and reinforcement training data. Sim-to-Real Reinforcement Learning and World Models (9:46) Unitree trains dynamic locomotion policies in physics simulators—including NVIDIA Isaac Sim, MuJoCo, and Gazebo—before deploying them directly to hardware. Public demonstrations, including televised group performances at China's CCTV Spring Festival Gala in 2024 and 2025, catalyzed national interest and contributed to the establishment of more than 100 humanoid robotics startups in China within a year. Technical benchmarks achieved through these reinforcement learning pipelines include a 3-meter aerial flip, continuous backflips, and a continuous freestyle table tennis rally. To broaden practical deployments, Unitree introduced 'Uni-World,' an open-source world model aimed at bridging the sim-to-real divide to automate tasks such as laundry, dishwashing, healthcare support, and warehouse sorting. Building an Ecosystem as a Competitive Moat (13:36) Facing accelerating domestic and international competition, Xu asserts that mechanical specifications are insufficient to maintain a defensible market position. Unitree has instead focused on embedding its hardware across academic and commercial research pipelines, with over 200 published scientific papers utilizing Unitree platforms and enterprise collaborations established with firms such as NVIDIA and Scale AI. Drawing parallels to ecosystem-centric technology firms like Apple and NVIDIA, Unitree's core strategy centers on pairing lower-cost, mass-produced hardware with open-source tools. By lowering financial barriers to hardware adoption, the company seeks to create long-term developer lock-in as research institutions and software developers standardize on its platforms. Unitree's roadmap relies on converting low-cost hardware distribution and simulation-trained models into an enduring platform moat; however, realizing its vision of self-manufacturing automation and commercial utility depends on how effectively developers can translate these platforms into reliable, real-world industrial and service applications.

精炼阅读摘要

宇树科技（Unitree Robotics）大客户负责人Nick Xu指出，公司正从四足机器人向涵盖紧凑型、全尺寸及轮式底盘的人形机器人矩阵全面演进。他认为，在生成式人工智能浪潮下，依托高性价比硬件、仿真强化学习与开源开发者生态构筑类似苹果与英伟达的产业护城河，是应对全球激烈竞争的核心路径。 从四足底盘到双足与轮式混合矩阵 (0:04) 宇树科技最初通过在教育、工业和消费领域交付四足机器狗建立全球市场认知。随着ChatGPT带动生成式人工智能热潮，以及行业对高可靠性、合理定价的科研平台需求激增，宇树于2023年推出了首款全尺寸双足人形机器人H1，并于2024年和2025年相继扩展出紧凑型G1和极具成本优势的R1机型。 针对仓储、产线和商业服务等对移动稳定性要求高于双足敏捷性的特定场景，宇树开辟了轮式混合产品线（如G1D、R1D与H2D），将双足机器人的上半身躯干集成至自动导引车（AGV）底盘上。与此同时，其全尺寸产品线也持续演进，推出了升级版本的H2和H2+机型。 规模化量产与自制造自动化愿景 (2:34) 紧凑型G1在2024年5月的IEEE机器人与自动化国际会议（ICRA）亮相后，迅速从科研测试平台转变为出货主力。据介绍，G1仅在2025年的全球销量就突破了5500台，推动其累计生产总量超过11000台。为了进一步降低人形机器人的普及门槛，宇树推出了起售价约4900美元的R1机型，该设计灵感源自钢铁侠，并入选了《时代》周刊2025年最佳发明榜单。 在生产制造端，宇树提出了利用人形机器人组装自身机器人的全自动化远景。工厂实录展示了G1机器人在实际产线上进行精密分拣并组装电机模块，这类复杂操作技能主要依赖人类操作员遥操作所收集的示范数据进行端到端学习。此外，轮式G1D还被作为具身智能的高效数据采集平台，支持将训练数据回传至本地私有云或AWS等公有云架构。 工业级硬件配置与灵巧手协同生态 (5:56) 面向重负荷工业自动化与高端科研需求，全尺寸H2人形机器人拥有31个自由度（DoF），整机重量约为150磅，计算单元可选配英特尔酷睿i7处理器与英伟达Jetson AGX Orin模块。宇树还与英伟达联合开发了H2+参考平台，该平台专门适配了雄克（Schunk）工业级灵巧手，以服务高级机器人开发者。 在末端执行器生态上，宇树采取了自研与开放兼容并行的策略。除自研的Dex5（适配H2）、Dex3-1（跨机型适配R1、G1与H2）和Dex1夹爪外，其机器人硬件还广泛支持无缝集成包括欧及（Ouji）、Robotiq、因时机器人（Inspire-Robots）以及雄克等第三方主流厂商的灵巧手组件。 强化学习突破与开源世界模型应用 (9:46) 宇树机器人的运动控制策略主要在英伟达Isaac Sim、MuJoCo以及Gazebo等物理仿真环境中通过强化学习进行训练，随后完成从虚拟到实体的部署迁移。在公共文化层面，六台H1机器人在2024年央视春晚的群舞表演获得了广泛关注，Nick提到该事件间接促成了国内百余家人形机器人初创公司的涌现；2025年春晚亮相的“武博（WuBot）”则展示了更进一步的动态控制能力。 目前，宇树机器人已实现了包括全尺寸人形机器人连续自由式乒乓球对打、3米高空翻以及连续后空翻等多项动态机动纪录。在具身大模型方面，宇树推出了开源的“Uni-World”世界模型，通过缩减仿真与物理世界的鸿沟，推动机器人在洗碗、折叠衣物、医疗陪护、超市理货及农业采摘等复杂真实场景中执行全自主作业。 借鉴苹果与英伟达模式构筑生态护城河 (13:36) 宇树目前的全球合作伙伴包括英伟达、Scale AI以及全球数百所顶尖高校与科研机构，基于宇树机器人硬件平台发表的学术论文已超过200篇。面对国内外日益激烈的市场竞争，Nick强调单纯依靠硬件参数与规格并不足以确立长期领导地位。 宇树的目标是借鉴苹果和英伟达的平台模式，将开放且具有粘性的生态系统作为核心竞争壁垒。通过持续提供高可靠性、高性价比的硬件底座，并配合开源算法与世界模型，宇树希望让全球开发者与企业客户长期深度绑定在其生态系统之内。 Nick Xu总结认为，人形机器人产业的竞争正从单一硬件性能比拼转向生态粘性构建。尽管宇树在出货规模、低成本控制以及仿真强化学习方面建立了先发优势，但在海内外大量初创企业与科技巨头加速涌入的行业格局下，如何将自组装产线落地成熟，并持续推动开源世界模型在实体工业场景中的规模化商业变现，仍是其生态战略面临的关键考验。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2026 年 8 月 21 日发布在 YouTube 平台上。视频简介显示，宇树科技（Unitree Robotics）大客户业务负责人 Nick Xu 在演讲中系统介绍了宇树的人形机器人全线产品矩阵及未来生态路线图。 ## TL;DR 演讲全面回顾了宇树从四足机器人跨越至双足及轮式人形机器人全矩阵的发展历程，重点剖析了突破万台量产的明星科研平台 G1、全尺寸旗舰 H2/H2+ 以及高性价比机型 R1。现场展示了 G1 在流水线组装自身电机的“机器人制造机器人”实测，以及在春晚舞台凭借强化学习创下的跳桌与高空抛掷世界纪录。此外，宇树公布了 UniForm 开源世界模型的多场景自主作业进展，并指出面对激烈竞争，公司正全力构建类似 Apple 与 NVIDIA 的全球开发者生态护城河。 ## Atmosphere & style 整个演讲节奏明快、务实干练，充盈着硬核科技的自信与突破感。讲者将惊艳眼球的高难度武术与空翻真机演示，同沉甸甸的量产出货数据与底层算法突破紧密结合，既有前沿科技破圈的激情，又透露出清晰冷静的商业生态战略思考。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) 宇树科技从四足机器人起家，受生成式 AI 爆发及具身智能硬件需求驱动，于 2023 年推出全尺寸双足人形机器人 H1 正式进军该赛道。 - [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) 宇树已构建完善的双足人形机器人梯队（H1、H2/H2+、G1、R1）及结合 AGV 底盘的轮式系列（G1D、R1D、H2D）。 - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) G1 通过物理仿真强化学习和 Sim-to-Real 技术实现了侧空翻、打功夫等高动态动作，并支持加装 Jetson AGX Thor 算力背包。 - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) G1 人形机器人实现了高度成熟的工业化量产，2025 年全球出货量超 5,500 台，累计生产制造量已突破 11,000 台。 - [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) 轮式人形机器人 G1D 专为数据采集与模型训练设计，在装配、仓储、巡检及商超等平整地面场景中具备更优的稳定性和续航表现。 - [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) 全尺寸旗舰 H2 拥有 31 个自由度，宇树联合 NVIDIA 推出了基于 H2 的 H2+ 官方参考设计平台。 - [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) 小型机型 R1 起售价仅约 4,900 美元并入选《时代》2025最佳发明；末端支持 Dex 系列自研灵巧手及帕西尼、BrainCo、Shadow Hand 等第三方集成。 - [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) 宇树工厂已落地 G1 机器人组装自身电机零部件，长远目标是实现整条生产线的完全无人化与“机器人制造机器人”。 - [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) 春晚武术节目《武魂》（WuBot）创下全球首例自由式跳桌/跳马、3 米高空抛掷空翻平稳落地等多项世界纪录，验证了仿真训练算法的跨越。 - [12:05](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=725s) 宇树正式开源 UniForm 世界模型，并在 G1 上验证了居家养老递药、洗碗机操作、洗衣折叠和商超货品分拣等全自主作业能力。 - [12:05](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=725s) 面对激烈的行业竞争，宇树致力于打造类似 Apple 和 NVIDIA 的繁荣生态护城河，全球基于其平台发表的顶会论文已超 200 篇。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) 宇树科技人形机器人演进与全矩阵产品布局 宇树科技（Unitree Robotics）的大客户业务负责人 Nick Xu 登台介绍了公司在构建下一代人形机器人及其生态系统方面的最新进展。回顾宇树的发展历程，公司起初是以四足机器人（机器狗）闻名，产品广泛销往教育科研、工业巡检及大众消费市场。随着以 ChatGPT 为代表的生成式 AI 掀起全球技术浪潮，加上学术界与工业界客户急需一款性能可靠、价格亲民的具身智能科研硬件平台，宇树于 2023 年顺势推出了第一款全尺寸双足人形机器人 H1，正式进军人形机器人赛道。 随后几年内，宇树迅速完善了产品梯队：2024 年推出了小尺寸人形机器人 G1，2025 年推出了极致性价比机型 R1；在全尺寸人形机器人领域，宇树在 H1 的基础上迭代升级推出了 H2 以及今年与合作伙伴联合打造的 H2 Plus。为了适应多样化的工业与商业场景，宇树还将双足人形机器人的上半身与 AGV 移动底盘相结合，衍生出了全系列的轮式人形机器人（Wheeled Humanoid），包括 G1D、R1D 和 H2D。 ```mermaid flowchart TD accTitle: 宇树科技人形机器人产品矩阵 subgraph 双足人形系列["双足人形机器人 (Bipedal)"] H1["H1 (2023 全尺寸初代)"] --> H2["H2 / H2+ (全尺寸旗舰 31自由度)"] G1["G1 (2024 经典科研/量产旗舰)"] R1["R1 (2025 入门级/高性价比)"] end subgraph 轮式人形系列["轮式人形机器人 (AGV 底盘)"] G1D["G1D (数据采集与训练)"] R1D["R1D (商用服务/轻量)"] H2D["H2D (工业自动化/科研)"] end ``` ### [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) 现象级明星产品 G1 与 G1D 轮式数据采集平台 在宇树的人形机器人矩阵中，G1 无疑是最具代表性且销量极高的经典标杆。G1 于 2024 年 5 月在机器人学术顶级会议 ICRA 上首次发布，最初定位为纯粹的学术科研平台。随着强化学习（Reinforcement Learning）算法的突破，宇树团队利用 NVIDIA 仿真环境以及 MuJoCo、Gazebo 等物理仿真平台进行大规模运动控制（Locomotion）训练，再通过 Sim-to-Real 技术迁移部署到真机上，使得 G1 能够稳定完成侧空翻、跳舞、打中国功夫等高难度动态动作。在硬件设计上，G1 头部集成了 3D 激光雷达（LiDAR）与深度相机，机身标配或可选配 NVIDIA Jetson Orin NX 计算单元，甚至支持在背部加装更强劲的 Jetson AGX Thor 算力背包。 为了解决具身智能领域对高质量现实数据采集的迫切需求，宇树推出了轮式版本的 G1D。G1D 专为机器人数据收集与模型训练而设计，采集到的交互数据既可以上传至 AWS 等公有云，也可存储在用户自建的本地私有云中。在装配车间、仓储物流、能源巡检、餐饮服务及大型超市等平整地面场景下，轮式底盘在移动稳定性和续航能力上显著优于双足步态，能够以更低成本高效落地作业。 - **G1 销量与产能数据**：仅在 2025 年一年内，G1 在全球范围内的出货量就突破了 5,500 台；截至演讲时，累计生产制造量已超过 11,000 台，建立了高度成熟稳定的工业化量产能力。 - **G1D 核心应用场景**：覆盖精密组装、自动化仓储拣选、商超货物整理、餐饮与能源巡检服务等。 --- ### [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) 全尺寸旗舰 H2、极致性价比 R1 与丰富末端灵巧手生态 在全尺寸领域，宇树带来了重量级产品 H2。H2 拥有 31 个自由度（DoF），整机重量约为 150 磅（约 68 kg），机载算力支持 Intel i7、NVIDIA Jetson Orin NX 以及最新的 Jetson AGX Thor 等多种计算模组组合。基于 H2，宇树今年还联合 NVIDIA 共同发布了 H2+ 作为面向全球开发者的全新参考设计平台（Reference Platform），其末端搭配了高性能灵巧手（如 Shadow Hand），为高阶具身智能算法研发提供顶级软硬件载体。此外，对应的轮式版本 H2D 则聚焦于工业自动化研发、高校前沿教学及高端商业服务。 在小型化与大众化方向上，宇树推出了外形颇具“钢铁侠”风格的小型人形机器人 R1。R1 拥有极高的性价比，起售价仅为 4,900 美元左右，成功入选了《时代》（*TIME*）杂志评选的“2025 年度最佳发明”（Best Inventions of 2025）。在末端执行器（灵巧手）配置方面，宇树构建了高度开放与自研兼备的生态体系： - **自研末端系列**：包括适配 H2 的 Dex 5 五指灵巧手、通配于 R1/G1/H2 的 Dex 3-1 三指灵巧手，以及工业级 Dex 1 双指夹爪。 - **第三方灵巧手兼容集成**：系统全面支持无缝集成包括触觉灵巧手（如帕西尼/无及科技）、BrainCo（强脑科技）、因时机器人（Inspire-Robots）以及 Shadow Hand 等业界主流第三方硬件。 | 型号 | 定位与形态 | 关键参数 / 硬件特性 | 算力与末端配置 | 典型应用领域 | |:--- |:--- |:--- |:--- |:--- | | **G1 / G1D** | 经典双足 / 轮式数据平台 | 3D 激光雷达 + 深度相机；累计量产超 1.1 万台 | Jetson Orin NX / AGX Thor 背包；Dex 3-1 / 夹爪 | 具身智能科研、数据采集、商超/仓储服务 | | **H2 / H2+ / H2D** | 全尺寸旗舰双足 / 轮式 | 31 自由度，重约 150 磅；H2+ 为 NVIDIA 官方参考平台 | Intel i7 / AGX Thor；Dex 5 / Shadow Hand | 工业组装、高端科研自动化、商业服务 | | **R1 / R1D** | 小型入门双足 / 轮式 | 酷似钢铁侠外观；荣获《时代》2025最佳发明 | 起售价约 $4,900；兼容 Dex 3-1 及多款灵巧手 | 极客开发、高校教育、消费级具身智能探索 | --- ### [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) “机器人制造机器人”工厂实测与春晚武术破纪录展示 Nick Xu 展示了一段宇树工厂内部的实机测试视频，画面中 G1 人形机器人正在流水线上精准组装它自己的电机零部件。宇树长远的终极目标是通过人形机器人来生产和制造更多的人形机器人，逐步实现整条制造生产线的完全无人化与自动化。目前工厂通过遥操作（Teleoperation）技术由人类工程师进行动作示教，在实时作业的同时源源不断地采集轨迹数据用于算法模仿学习与泛化训练。 谈及行业影响力，宇树的人形机器人在国内乃至全球多次创造里程碑事件： - **初代春晚群舞破圈**：在上一届央视春晚上，宇树部署了 6 台全尺寸 H1 人形机器人完成了高难度排舞表演，在全网斩获数以亿计的播放量。这次公开亮相直接点燃了国内人形机器人产业的投资与创业狂潮，仅在随后一年内国内便涌现出上百家人形机器人初创企业，极大加速了整个行业的工业化进程。 - **最新春晚《武魂》（WuBot）武术绝技与世界纪录**：在最新的春晚舞台上，宇树呈现了名为《武魂》（WuBot）的人形武术群控表演，集中展现了一年来的算法跃迁，创下了多项世界纪录： - **全球首例连续自由式桌上腾空跳跃/跳马动作**。 - **全球首例 3 米高空抛掷空翻落地受控平稳站立**。 - **多机连续自主后空翻**。 这些超高难度动作完全依托于强化学习算法在仿真环境中历经数千万次迭代训练生成，随着 Sim-to-Real 技术鸿沟不断缩小，真机落地执行极端动作的成功率已大幅攀升。 --- ### [12:05](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=725s) UniForm 开源世界模型与构建类苹果/英伟达的生态护城河 在软件与具身智能模型层面，宇树正式开源了其自主研发的 UniForm 世界模型（UniForm World Model），并利用 G1 机器人在多类真实物理场景中展开全自主作业训练。演示展现了 G1 在居家养老陪护中精准为老人递送药物、在厨房内全自主抓取碗碟放入洗碗机清洗、熟练操作洗衣机清洗折叠衣物，以及在商超和农场仓库中进行货品分拣与货架整理。未来，这些能力将全面迁移并部署到全尺寸的 H2 平台上进行更大规模的泛化测试。 ```mermaid flowchart LR accTitle: 宇树具身智能世界模型与生态飞轮 subgraph 算法训练与世界模型["UniForm 开源世界模型"] SIM["仿真环境大规模强化学习"] --> S2R["Sim-to-Real 缩小现实差距"] S2R --> TASK["自主家政/医疗/商超/仓储作业"] end subgraph 全球开发者生态["开发者与客户生态护城河"] HW["高性价比稳定硬件 (G1 / H2 / R1)"] --> ECO["全球高校 / NVIDIA / Skild AI 等伙伴"] ECO --> PAPERS["200+ 顶会学术论文产出"] PAPERS --> RET["强客户粘性与开发者生态锁定"] end TASK -.闭环数据反馈.-> HW ``` Nick Xu 总结道，当前尤其是在中国市场，人形机器人领域的同质化竞争已变得异常激烈。宇树认为，企业真正的核心竞争护城河绝非单一的硬件参数，而是建立像 Apple 和 NVIDIA 那样繁荣闭环的生态系统（Ecosystem）。目前包括 NVIDIA、Skild AI 以及全球顶尖高校在内的众多机构都在采购 G1 作为基准平台，全球基于宇树人形机器人发表的顶会学术论文已突破 200 篇。只要能够持续迭代交付极具产品力的硬件底座，让全球开发者和商业客户在宇树的生态圈内持续沉淀代码、算法与应用，这种开发者生态与平台黏性便会成为宇树最坚固的壁垒。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=4s) 学术界与工业界客户急需一款性能可靠、价格亲民的具身智能科研硬件平台，宇树于 2023 年顺势推出了第一款全尺寸双足人形机器人 H1，正式进军人形机器人赛道。 — Nick Xu > [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) 仅在 2025 年一年内，G1 在全球范围内的出货量就突破了 5,500 台；截至演讲时，累计生产制造量已超过 11,000 台，建立了高度成熟稳定的工业化量产能力。 — Nick Xu > [02:34](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=154s) 在装配车间、仓储物流、能源巡检、餐饮服务及大型超市等平整地面场景下，轮式底盘在移动稳定性和续航能力上显著优于双足步态，能够以更低成本高效落地作业。 — Nick Xu > [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) 基于 H2，宇树今年还联合 NVIDIA 共同发布了 H2+ 作为面向全球开发者的全新参考设计平台（Reference Platform）。 — Nick Xu > [05:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=356s) R1 拥有极高的性价比，起售价仅为 4,900 美元左右，成功入选了《时代》（TIME）杂志评选的“2025 年度最佳发明”。 — Nick Xu > [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) 宇树长远的终极目标是通过人形机器人来生产和制造更多的人形机器人，逐步实现整条制造生产线的完全无人化与自动化。 — Nick Xu > [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) 这次公开亮相直接点燃了国内人形机器人产业的投资与创业狂潮，仅在随后一年内国内便涌现出上百家人形机器人初创企业，极大加速了整个行业的工业化进程。 — Nick Xu > [08:56](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=536s) 这些超高难度动作完全依托于强化学习算法在仿真环境中历经数千万次迭代训练生成，随着 Sim-to-Real 技术鸿沟不断缩小，真机落地执行极端动作的成功率已大幅攀升。 — Nick Xu > [12:05](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=725s) 宇树认为，企业真正的核心竞争护城河绝非单一的硬件参数，而是建立像 Apple 和 NVIDIA 那样繁荣闭环的生态系统（Ecosystem）。 — Nick Xu > [12:05](https://www.youtube.com/watch?v=7mZvwFtF3vU&t=725s) 只要能够持续迭代交付极具产品力的硬件底座，让全球开发者和商业客户在宇树的生态圈内持续沉淀代码、算法与应用，这种开发者生态与平台黏性便会成为宇树最坚固的壁垒。 — Nick Xu ## Mentioned Nick Xu, 宇树科技 (Unitree Robotics), Ai4, H1, H2, H2+, H2D, G1, G1D, R1, R1D, Dex 5, Dex 3-1, Dex 1, Shadow Hand, BrainCo (强脑科技), 因时机器人 (Inspire-Robots), 帕西尼感知科技, NVIDIA (英伟达), Jetson Orin NX, Jetson AGX Thor, MuJoCo, Gazebo, AWS, Intel i7, 《时代》杂志 (TIME), ICRA, 央视春晚, 《武魂》(WuBot), UniForm 世界模型, Apple (苹果), Skild AI, ChatGPT ## Video description Nick Xu - Key Account Lead, Unitree Robotics I will introduce about Unitree's Product and future roadmap. [Source](https://www.youtube.com/watch?v=7mZvwFtF3vU)

Source quotes

[0:04] Nick Xu: 学术界与工业界客户急需一款性能可靠、价格亲民的具身智能科研硬件平台，宇树于 2023 年顺势推出了第一款全尺寸双足人形机器人 H1，正式进军人形机器人赛道。 [2:34] Nick Xu: 仅在 2025 年一年内，G1 在全球范围内的出货量就突破了 5,500 台；截至演讲时，累计生产制造量已超过 11,000 台，建立了高度成熟稳定的工业化量产能力。 [2:34] Nick Xu: 在装配车间、仓储物流、能源巡检、餐饮服务及大型超市等平整地面场景下，轮式底盘在移动稳定性和续航能力上显著优于双足步态，能够以更低成本高效落地作业。 [5:56] Nick Xu: 基于 H2，宇树今年还联合 NVIDIA 共同发布了 H2+ 作为面向全球开发者的全新参考设计平台（Reference Platform）。 [5:56] Nick Xu: R1 拥有极高的性价比，起售价仅为 4,900 美元左右，成功入选了《时代》（TIME）杂志评选的“2025 年度最佳发明”。 [8:56] Nick Xu: 宇树长远的终极目标是通过人形机器人来生产和制造更多的人形机器人，逐步实现整条制造生产线的完全无人化与自动化。 [8:56] Nick Xu: 这次公开亮相直接点燃了国内人形机器人产业的投资与创业狂潮，仅在随后一年内国内便涌现出上百家人形机器人初创企业，极大加速了整个行业的工业化进程。 [8:56] Nick Xu: 这些超高难度动作完全依托于强化学习算法在仿真环境中历经数千万次迭代训练生成，随着 Sim-to-Real 技术鸿沟不断缩小，真机落地执行极端动作的成功率已大幅攀升。 [12:05] Nick Xu: 宇树认为，企业真正的核心竞争护城河绝非单一的硬件参数，而是建立像 Apple 和 NVIDIA 那样繁荣闭环的生态系统（Ecosystem）。 [12:05] Nick Xu: 只要能够持续迭代交付极具产品力的硬件底座，让全球开发者和商业客户在宇树的生态圈内持续沉淀代码、算法与应用，这种开发者生态与平台黏性便会成为宇树最坚固的壁垒。

Source quotes

[7:26] Nick Xu: the most cost-effective humanoid robot

Original transcript

[0:04] Um hello everyone. I'm Nick [0:07] uh from Uni Robotics. So it's glad to  be here and share the presentation. So [0:15] um today I'm going to introduce is the building  the next generation humanoid robot and ecosystem. [0:25] Um so first I want to show you the product  matrix of unitry. So uh as you can see uh at [0:34] first we just uh the robot dog or quad robot  manufacturer right. So um we are we have sold [0:43] a lot to educational market industrial market and  uh also the consumer market. And as for the human [0:53] robot actually uh you can see this is our first  model H1 humanoid robot we released in 2023 and [1:05] uh the reason we uh produced the human robot uh  is that uh I think first reason is that the AI of [1:15] chap GBT is become a huge trend and also some of  our customers they uh some demand who needs some [1:25] um reliable and also uh reasonable with reasonable  price platform to do their research. So that's the [1:35] main reason we start enter into the human robot  and uh in 2024 we released our uh G1 human robot [1:49] and also in 2025 we have released the R1 and for  the full size as you can see we've released the H2 [2:00] which is an upgrade version based on H1 one and H2  Plus this year. And we also have some models for [2:10] the wheelbase. We use half like up body of our U  bipedal human robot and mount it on the on AGV. So [2:21] it's a wheelbase robot. So we have G1D, R1D, and  H2D here. Yeah. So next I'm going to introduce our [2:34] most successful version G1. So G1 will release in  uh 2024 in May during academic conference called [2:46] Ikra and uh is right now is the most I think it's  most classic and bestselling robot until now. [2:56] uh we in last last year 2025 we have sold more  than 5,500 units globally and until now we have [3:08] produced more than 11,000 units. Yeah. So  I think you can see probably you can see [3:15] all kinds of videos online G1 doing some dancing  or kung fu. So all kinds of fancy movements. So [3:26] uh actually at first we released G1 uh it's just  purely a platform for research purpose. So but [3:34] with advancing of AI and uh the reforms learning  we have tried a lot with you know the local motion [3:43] side. So we chain it in the Nvidia simulation  software and some like Mojo Gazbo this kind of [3:52] software and uh we deployed on our robots that's  how we realize all these fancy movements like [4:00] uh flip like dancing. Yeah. And to also  with a very good stability. Yeah. So [4:09] uh you can see this is the design is actually  quite good like we it has two sensor like the [4:17] 3D lighter and depth camera on the head and also  as for the uh GPU we have optional Nvidia JSON or [4:28] ring NX and even a backpack like AGX store can be  mounted on the back. Yeah. So we have different [4:37] kind models we keep updating it. We have very  stable production capability for G1 right now. And [4:47] uh as for the G1D release uh last year so this  is should be a platform for cumula robot data [4:55] and training. So we have a platform which you can  use the G1D the wheelbase robot to collect data [5:04] and also as for data you collect you can choose  to mount on the public cloud like AWS or your [5:14] personal local cloud. Yeah. So uh we like I think  it's more suitable for a lot of scenarios like [5:24] uh assembly uh warehouse and energy or beverage  service and supermarkets. So wheelbased I think uh [5:35] the advantage is that the stability is better than  the the leg robot. So on under these scenarios, [5:45] I think they're better than the U by pedal him  robot. Uh the next version I'm going to introduce [5:56] is our H2 him robot. So um this is a with 31  degree freedom and 150 pounds which is quite heavy [6:07] obviously. And um also as for the compute power it  has we also have uh some option like uh Intel i7 [6:19] uh JSON or X or JSON AGX store. So um yeah  so this is some option GPU options we have [6:31] uh as for the H2+ this is actually uh reference  platform we work together with Nvidia right so um [6:41] this is we you use the uh our human robot body as  for the tech hands is uh sharper hand So uh this [6:54] is uh yeah a new version released this year and uh  this is a reference platform for all the develop [7:02] developers and um the next is the the G1D uh the  wheelbased uh sorry H2D and uh this is H2 which is [7:14] suitable some for some uh industrial automation  research and education and commercial service. [7:26] Uh the next one I'm uh going to introduce is our  R1. So this is a small size human robot and uh I [7:37] think is the most cost effective human robot. The  price is only around like start from 4,900 which [7:45] uh yeah is quite cheap and uh also uh the design  is uh well somehow looks like an Iron Man. Yeah. [7:56] And uh our has achieved like uh from time magazine  where like the the best inventions of 2025. [8:10] Yeah. So the next is also same as the wheelbase.  We use the hard body of our model on the on a AGB. [8:22] So this is some hand options we have for  our human robot like uh we have our own [8:28] Dex 5 hands can be mounted on H2 Dex 3-1  can be m mounted on all R1 G1 H2 and Dex [8:39] one gripper as well and we can also integrate  with some third party hands like Uji hand like [8:47] uh Brinko hands inspire hands and sharp our hands. [8:56] So, uh this is a video that we are  doing some testing our own factory [9:01] where training the G1 to assemble the mode  its own moto. So that's like our final goal [9:12] is will be using the human robot to you  know produce more robots. Yeah. So we're [9:19] trying to automate our production line with  our own robots. So you can can see doing some [9:26] uh sorting tasks and also the assembly tasks.  So this is training by teleoperation. Uh we [9:37] have a some operator teleoperated robot. So  at the same time and doing doing the training. [9:46] So uh this is some of I think will be the the  milestone for the human robot industry in China. [9:57] So during the uh spring festival gala of last  year, we're using six H1 robots to do the group [10:06] dancing. Yeah. I think we got like millions  of views for this uh stage performance. Yeah. [10:14] So I think uh after this you know spring festival  gala the whole human robot industry in China like [10:24] progress a lot there's like more than 100 human  robot company were funded in China last year so [10:35] yeah I think we're we kind of uh accelerate  the process for the whole industry Okay. [10:46] So actually this year we did also uh went to  the spring festival gala as well. Uh the show [10:55] is called wubot. So it's kind of refle reflect the  progress we achieved within the one you know past [11:04] one year. So it actually we achieved several  world records. For example, it's world first [11:13] continuous freestyle uh table fing pu and also uh  world's first launched area flip. The height is [11:23] like 3 mters height and also you can see can do  multiple continuous flip. So all these actually [11:32] are achieved by reforms learning algorithm. You  use the simulation software to train thousand of [11:39] times and then from sim to real deploy on the  real robot. So the success rate there's a gap [11:46] for sure but all these locomotion actually  you can see the gap is becoming smaller and [11:52] smaller. So the success rate is becoming high  to achieve all these fancy you know movements. [12:05] So uh actually we also open sourced some word  models like this is our uniform word model you [12:14] can see we're using G1 to do a uh in some certain  scenarios to do the training like uh in house [12:23] care scenarios the m can serve the medicine for  too and like um yeah so this all this is fully [12:35] automous and doing some household tasks as well  like cleaning dishes in the you know dishwashing [12:46] machine and also in some supermarket scenarios  you can do do some sorting and organizing top. [13:03] So yeah, so now we're using a lot of G1 to do  the training. I think um very soon we'll use [13:09] our H2 to do the testing and chaining as well.  Yeah. So now this is uh doing laundry for you [13:24] and it also can work in some like warehouse and [13:29] farms as well. So this is  open sourced on our website [13:36] and um I think you know this is the  global ex ecosystem you can see lots [13:46] of companies are using our robots. Uh this  is like just some part of it like Nvidia, [13:54] Skilled AI and a lot of universities are using  the G1 for research and public a lot of lots of [14:01] papers. I think there are already more than maybe  200 papers have been released based on our robots. [14:08] So um yeah so you can see actually the  competition in China especially in China is [14:16] uh becoming you know a lot like very extreme but  you know um I think the the competitive modes [14:26] for the companies to build the the ecosystem. So  seem like Apple like u um Nvidia you can build the [14:36] ecosystem around you. You you can keep releasing  you know uh very good product and your customer [14:45] will keep using your product. So this I think  this will be like a key mode for for our company.