文 | 产联社CLS
“明年很大概率会是整个机器人行业或者机器人模型的一个临界点。”11月4日，宇树科技创始人、董事长兼首席执行官王兴兴在2026年世界互联网大会乌镇峰会“互联网企业家”分论坛上表示，当前全球顶"尖"人工智能模型已具备较强的理解与规划能力，但机器人在精细操作及与物理世界交互方面仍存在技术瓶颈。
他透露，宇树科技已开始探索物理人工智能模型的自主进化，让模型自行编写代码、训练和迭代，人类主要负责评价和设定目标。
展望未来，他认为，除数据和算力外，芯片产能以及背后的能源、矿产资源也将成为全球科技产业发展的关键资源。
2027年或迎产业临界点
王兴兴表示，过去几年，宇树科技持续推动机器人进入工厂等实际作业场景。目前，在部分单工序或固定工位上，经过充分训练的机器人模型作业成功率可以接近100%，但作业效率仍有进一步提升的空间。
相比固定场景下的作业能力，他更关注机器人的泛化能力，即机器人能否在不同环境中快速部署，并完成各种实际任务。
“我们希望做到的是，在任意场景、任何地方部署一台机器人，就能非常方便地干活。”王兴兴表示，2027年的技术发展窗口很大，机器人行业或机器人模型很可能迎来一个临界点。
他进一步提出，具身智能有望迎来“奇点时刻”：在约80%的共通场景中，机器人能够通过语音指令完成约80%的任务。他认为，全球顶尖模型距离这一目标已经非常接近。
不过，机器人要真正进入复杂的生产生活场景，仍面临精细操作能力不足的挑战。
王兴兴指出，目前全球头部多模态模型在空间、时间规划方面已经表现出较强能力，简单抓取和搬运任务的成功率也较高，但在手机拆装、穿针引线等需要精密操作的任务中，成功率仍然偏低。
“目前最大的问题就是，最后的几个毫米问题其实没有解决好。”他认为，现阶段人工智能模型的输入输出与真实物理世界之间仍存在差距，尤其是在接触、触觉等方面。
王兴兴将这一问题形象地比作机器人的“小脑”能力。他认为，如果未来一至三年能够解决模型与物理世界的对齐问题，物理人工智能模型面临的主要理论障碍将有望进一步减少。
内部已开启物理大模型自进化预研
除产品开发与场景落地外，王兴兴还介绍了宇树科技在物理人工智能模型方面的前沿预研方向。
他透露，宇树科技几个月前已开始在内部探索物理人工智能模型的自主进化，并在现场展示了“物理AI机器人自进化V1.0”的预研方案。
宇树科技提出四个方向：基础模型越强，自进化性能越强；多源数据规模化，驱动自进化规模效应；通过更多真机部署和数据累加，实现指数增长的自进化速率；持续累加扩展技能，走向规模化技能生态。
王兴兴表示，全球头部人工智能大模型企业已在内部建立较为成熟的自主进化循环。模型能够自行采集数据、开展训练并持续迭代，人类主要参与评价。相比模型架构本身的变化，模型自主采集数据、训练及强化学习能力的提升，正成为推动技术进步的重要因素。
不过，在机器人领域，包括宇树科技在内的全球机器人企业，对这一技术路径的应用仍相对有限。
按照王兴兴介绍的技术设想，人工智能模型可以自主编写代码，在仿真或真实环境中完成训练，再将训练结果部署到机器人上进行测试。系统根据测试成功率进行评价，并将结果反馈给模型，由模型自主修改代码、重新训练，形成持续迭代的闭环。
在这一过程中，人的角色也可能发生变化。“到现在为止，我觉得人只剩下一个最大价值就是打分。”王兴兴表示，在模型自主进化的过程中，人类主要负责评价模型表现、设定目标和判断结果是否符合预期。
他进一步指出，随着模型逐步学习人类的评价标准和偏好，部分原本需要人类参与的监督、评价工作，也可能由人工智能自主完成。
王兴兴以视频生成模型为例表示，早期视频生成需要人类大量参与导演、剪辑等环节，而随着模型能力提升，未来模型有望更加自主地理解人的偏好，并持续优化生成结果。
对于机器人企业而言，模型自主进化不仅意味着研发方式的改变，也可能提高数据利用效率。
王兴兴坦言，目前宇树科技在数据集利用方面仍存在不足，大量数据尚未被人工智能模型充分使用。他认为，随着基础模型的逻辑推理与自主迭代能力提升，机器人企业有望更充分地利用已有数据，加快技术研发。
芯片产能及其上游资源很关键
谈及人工智能时代的长期产业竞争，王兴兴将关注点从机器人产品和模型能力进一步延伸至算力基础设施及其上游资源。
他表示，近年来一直在思考一个问题：对于个人和企业而言，未来最重要的资源究竟是什么？
在他看来，数据和算力固然重要，但要支撑这些资源持续发挥作用，芯片产能仍是全球范围内非常稀缺的关键资源。
王兴兴表示，当前国内算力资源仍然紧缺，而芯片生产背后还涉及能源、铜等矿产资源，这些上游资源同样具有重要的战略价值。
他以英特尔为例指出，随着人工智能产业对算力需求持续增长，芯片制造企业未来可能需要建设更多大规模制造设施，其产业角色也可能进一步向算力生产环节延伸。
王兴兴认为，未来科技产业的发展不仅取决于人工智能模型本身的进步，还将受到芯片制造能力、能源供给及关键矿产资源等基础条件的影响。
近半数产品出口海外
在全球市场方面，王兴兴介绍，宇树科技自2018年开始拓展海外业务，过去几年平均每年有接近50%的产品出口海外。
海外市场不仅是宇树科技的重要销售市场，也是其机器人技术开发和应用探索的重要参与者。
王兴兴透露，在宇树科技10月22日举办的物理人工智能合作伙伴大会上，现场约有3000名参会者，其中超过1000人来自海外，占比至少三分之一。
他介绍，全球不少客户和合作伙伴正基于宇树科技的机器人产品进行二次开发，探索建筑、商业服务、医疗及物流搬运等不同应用场景。
例如，美国研究人员利用宇树科技G1人形机器人完成活体胆囊切除，并发表了《自然》期刊论文，成为全球首例使用人形机器人完成的活体手术。日本羽田机场已在探索利用相关机器人开展简单搬运作业。
在产品方面，宇树科技近年来持续推出四足机器人、人形机器人以及灵巧手等产品。其中，新一代灵巧手拥有22个自由度，并在关节内部设置力矩限制器，以降低受到外部冲击时的损坏风险。
王兴兴还介绍，宇树科技近期发布了身高约1.7米的全尺寸人形机器人，并探索将其应用于安保等场景。他认为，海外市场在相关应用方面具有较为旺盛的需求。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
CLS
"The great probability of next year will be a critical point for the entire robot industry or robotic models." On 4 November, Wang Hing-Hing, founder, chairman and chief executive officer of Woo Tree Science and Technology, stated at the sub-forum “Internet Entrepreneurs” of the World Internet Assembly’s U-town Summit, 2026, that the top global “top-line” artificial intelligence model now has a strong understanding and planning capability, but that robots still have technical bottlenecks in precision and interaction with the physical world.
He revealed that Wootree technology had begun to explore the autonomous evolution of physical artificial intelligence models, allowing models to develop their own codes, training and trajectories, with humans primarily responsible for evaluating and setting targets.
Looking ahead, he believed that, in addition to data and data, the production of chips and the energy and mineral resources behind them would be key resources for the development of the global science and technology industry.
2027 or industry threshold
Wang Hing-Hing indicated that over the past few years, Woo-tree technology has continued to drive robotics into operational scenarios such as factories. At present, fully trained robotic model operations success rate can be close to 100 per cent on some individual processes or fixed positions, but there is room for further improvement in operational efficiency.
He was more concerned about robotics ' generalization capacity, i.e. their ability to deploy rapidly in different environments and to perform a variety of practical tasks, than was the ability to operate under a fixed scenario.
“What we want to do is to deploy a robot in any setting or anywhere, and work very easily.” Wang Xing Xing indicated that the technological development window in 2027 was large and that the robot industry or robotic models were likely to reach a critical point.
He further suggested that smart minds were expected to bring about “unusual moments”: in about 80 per cent of shared scenes, robots were able to perform about 80 per cent of their tasks through voice commands. In his view, the goal of a global top model was very close.
However, robots still face the challenge of being insufficiently sophisticated to actually enter complex productive life scenarios.
Wang Xing Xing noted that the current global head model has shown greater capacity in space and time planning, and that success rate is higher for simple capture and removal tasks, but that success rate remains low for tasks that require sophisticated operation, such as cell phone dismantling and needle lead.
“The biggest problem at the moment is that the last few millimeters have not actually been solved.” In his view, there was still a gap at the current stage between the input and output of artificial intelligence models and the real physical world, especially in terms of exposure, touch, etc.
Wang Xing Xing is imaged as a robotic “brain” capability. In his view, the main theoretical barriers to physical artificial intelligence models were expected to be further reduced if the alignment of models with the physical world was addressed over the next one to three years.
Internal LLM self-evolutional studies have been initiated.
In addition to product development and landscape development, Wang Hing-Hing also presented the forward-looking directions of WOIT in physical artificial intelligence models.
He revealed that Woo Tree technology had begun to explore the autonomous evolution of physical artificial intelligence models internally a few months ago, and that a pre-study programme for “Physico-AI robots self-evolution V1.0” had been shown on site.
Woo-tree technology proposes four directions: the stronger the underlying model, the stronger the self-evolutionary; the size of multi-source data, which drives the self-evolutionary scale effect; the rate of self-evolution for index growth through more real-time deployment and data accumulation; and the continuous build-up of skills to scale skills ecology.
Wang Hing Xing indicated that the global AAQ Z43Z enterprise has built a more mature autonomous evolutionary cycle within itself. Models allow for self-collection, training and continuity, with human participation in evaluations. Compared with changes in model structures themselves, the upgrading of model autonomous data collection, training and enhanced learning capabilities is becoming important drivers of technological progress.
However, in the robotic field, the application of this technology path remains relatively limited for global robotic enterprises, including ow tree technology.
According to the technology scenario presented by Wang Heung Heung, artificial intelligence models can create their own codes, complete their training in a simulation or real environment and deploy the results to robots for testing. The system is evaluated on the basis of testing success rate, and the results are fed back to the model, where the code is modified and retrained autonomously to form a closed circle of continuous overlap.
In the process, the role of the human person may also change. “To now, I feel that only one of the greatest values left is to score.” Wang Hing Xing indicated that in the process of autonomous evolution of models, human beings are primarily responsible for evaluating model performance, setting targets and judging whether the results are in line with expectations.
He further noted that, as models evolved to learn from human evaluation criteria and preferences, some of the monitoring, evaluation and possibly artificial intellectual autonomy that would otherwise require human participation could be achieved.
As an example of a video production model, Wang Xing Xing Xing indicated that early video generation required significant human participation in the director, clipping, etc., and that, as model capabilities increased, future models were expected to understand people ' s preferences more autonomously and continuously optimize the generation of results.
For robotic enterprises, the autonomous evolution of models means not only a change in the way research and development is developed, but also an increase in the efficiency of data use.
Wang Xin Xin Xin has stated that there are still gaps in the use of data sets in the field of u-tree technology and that much of the data is not yet fully used by artificial intelligence models. In his view, robotic enterprises were expected to make fuller use of available data to accelerate technological development as the logic of the underlying model and autonomous iterative capabilities increased.
Chip energy and upstream resources are critical.
Turning to long-term industrial competition in the age of artificial intelligence, Wang Xing Xing has further extended the focus from robotic products and modelling capabilities to computing infrastructure and its upstream resources.
He said that he had been thinking about a problem in recent years: What are the most important future resources for individuals and businesses?
In his view, while data and calculus were important, the production of chips remained a critical resource scarce globally to support the continued functioning of those resources.
Wang Xing Xing has indicated that domestic computing resources are still scarce and that mining resources, such as energy and copper, are also involved in the production of chips, which are also of strategic value.
Citing Intel as an example, he noted that, as the demand for computing in the artificial intelligence industry continued to grow, chip manufacturing enterprises might need to build more large-scale manufacturing facilities in the future, and their industrial role could be further extended to the computing chain.
Wang Xing Xing Xing believes that the future development of the science and technology industry will depend not only on advances in artificial intelligence models themselves, but also on the underlying conditions of chip manufacturing capacity, energy supply and key mineral resources.
Nearly half of the products are exported overseas.
With regard to the global market, Wang Xing Xing Hing described the expansion of WOST since 2018, with an average of nearly 50 per cent of products exported abroad each year over the past few years.
Overseas markets are an important market not only for the sale of tree technology but also for the development and exploitation of robotic technology.
Wang Hing Xing revealed that at the conference of physical artificial intelligence partners, held on 22 October by Woo Tree Science, there were some 3,000 participants on the scene, over 1,000 of whom were from abroad, representing at least one third of the total.
He described a number of clients and partners around the world who were redeveloping robotic products based on Wootree technology, exploring different applications such as construction, commercial services, medical care and logistics handling.
For example, the U.S. researchers, using the Waki technology G1 man-form robots to complete the biocyst removal and to publish thesis of Nature, became the first living surgery to be performed globally using human robots. Haneda Airport in Japan has explored the use of relevant robots for simple handling operations.
In terms of products, the technology has been continuously introducing four-foot robots, human robots and smart hands in recent years. Of these, a new generation of smart hands has 22 degrees of freedom and has internal force rectangles in their joints to reduce the risk of damage in the event of external shocks.
Wang Heung-Hing also described the recent development of publish in Woo-tree Technology of a full-sized human robot approximately 1.7 metres tall and the exploration of its application to security scenarios. In his view, overseas markets had a greater demand for relevant applications.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.