“我们现在看到机器人在跑步、武打、跳舞等方面的表现已经不错，但在复杂、精细作业方面还有很多技术问题需要解决。”西湖大学人工智能讲席教授、人工智能系主任金耀初在2026可持续全球领导者大会期间接受《中国经营报》记者采访时表示，当前市场对于具身智能部分技术进展和产业化速度的预期可能偏高。作为长期从事人工智能、类脑计算研究的学者，金耀初在2023年起担任西湖大学可信及通用人工智能实验室负责人。在他看来，具身智能真正走向产业应用，不能仅依赖大规模数据采集与模型训练，还需要在硬件、感知、控制以及自主学习能力等方面取得突破。
"We now see robots performing well in running, martial arts, and dancing, but there are still many technical problems to solve in complex, fine manipulation." Jin Yaochu, Chair Professor of Artificial Intelligence and Dean of the Department of Artificial Intelligence at Westlake University, told China Business Journal during the 2026 Sustainable Global Leaders Conference that current market expectations for some embodied intelligence technology progress and industrialization speed may be too high. A scholar long engaged in AI and brain-inspired computing research, Jin has headed Westlake University's Trustworthy and General AI Laboratory since 2023. In his view, embodied intelligence truly moving toward industrial application cannot rely solely on large-scale data collection and model training; breakthroughs are also needed in hardware, perception, control, and autonomous learning capability.
中国完整的制造业体系和丰富的应用场景能够为具身智能发展提供重要支撑。但具身智能从实验室走向真实应用仍需要时间，特别是在复杂工业操作、家庭服务等领域，技术成熟度与实际需求之间仍存在距离。今年以来，具身智能行业的融资热度持续上升。金耀初表示，中国具身智能领域的投资热度确实比较高，目前获得较大规模融资的具身智能企业已经有相当数量。但与此同时，大家也普遍存在一种疑问，就是这些技术到底能不能真正用起来。对于行业是否存在泡沫，我们需要更清楚地认识到，具身智能现在究竟能够做什么，未来又应该能做什么。
China's complete manufacturing system and rich application scenarios can provide important support for embodied intelligence development. But embodied intelligence still needs time to move from laboratory to real application, especially in complex industrial operations and home services, where distance remains between technology maturity and actual demand. Since the beginning of this year, embodied intelligence industry financing heat has continued rising. Jin said China's embodied intelligence investment heat is indeed relatively high, with a considerable number of embodied intelligence enterprises having obtained relatively large-scale financing. But at the same time, everyone generally has a question—can these technologies truly be used. Regarding whether the industry has a bubble, we need to more clearly recognize what embodied intelligence can do now and what it should be able to do in the future.
通过实际研究，我们意识到当前具身智能距离真正承担复杂任务，还有不少技术问题需要解决。例如，如果人形机器人真正进入产业应用，精细操作能力必须足够强。又例如，对于人来说，将两根柔性电线连接起来，再使用工具进行固定，可能是相对简单的工作，但对于机器人而言，完成这些操作具有很大的挑战性。目前不少具身智能研究还集中于桌面上的物体抓取、搬运等相对封闭的场景。这些研究有价值，但仅靠这些能力，还不足以支撑机器人完成更复杂的工作。所以，行业需要进行一些思考。不能只是增加企业数量，更重要的是让研发资源得到有效投入，使机器人真正具备解决实际问题的能力。
Through actual research, we realize current embodied intelligence still has many technical problems to solve before truly undertaking complex tasks. For example, if humanoid robots truly enter industrial application, fine manipulation capability must be sufficiently strong. Another example: for humans, connecting two flexible wires and then fixing them with a tool may be relatively simple work, but for robots, completing these operations is highly challenging. Currently much embodied intelligence research still concentrates on relatively closed scenarios like desktop object grasping and carrying. These studies are valuable, but such capabilities alone are insufficient to support robots completing more complex work. So the industry needs to think. It is not just about increasing enterprise numbers; more importantly, R&D resources must be effectively invested so robots truly possess the capability to solve practical problems.
在某些方面，市场对具身智能技术的期待可能确实比较高。具身智能研究非常重要，特别是与实际应用场景结合的研究。政府加大科研投入、支持校企合作、建设中试基地和训练基地，都有必要。但是，我们不能因此形成一种预期，认为明年机器人就能够广泛应用于各种复杂场景。技术的发展需要时间。我们当然希望具身智能能够发展得比较快，但它不会是今年或明年就能够完成的事情。因此，不太相信具身智能明年就会出现所谓的iPhone时刻。现在机器人在跑步、武打、跳舞等方面的表现确实不错，但精细作业能力仍然存在明显局限。尤其是进入家庭，可能还需要更长时间。比如让机器人照顾老年人，除了能不能完成任务，还要考虑安全性、可靠性等问题。
In some aspects, market expectations for embodied intelligence technology may indeed be too high. Embodied intelligence research is very important, especially research combined with actual application scenarios. Government increasing research investment, supporting school-enterprise cooperation, and building pilot bases and training grounds are all necessary. But we cannot therefore form an expectation that robots will be widely applied in various complex scenarios next year. Technology development needs time. We certainly hope embodied intelligence can develop relatively fast, but it will not be something completed this year or next year. Therefore, I do not quite believe embodied intelligence will see a so-called "iPhone moment" next year. Current robot performance in running, martial arts, and dancing is indeed good, but fine manipulation capability still has obvious limitations. Especially entering homes may require even longer. For example, letting robots care for the elderly—beyond whether tasks can be completed, safety and reliability must also be considered.
当前具身智能最主要的技术路线，还是端到端的数据驱动方法。数据采集是有必要的，但哪一种数据采集方式最有效，现在还很难判断。实验室也在采用不同方式开展研究，包括动作捕捉、VR设备等。不同采集方式都有各自的优点和局限性。比如，有些方式对于采集手臂和身体关节的运动比较有效，但是对于手指的精细动作，采集效果可能就不够理想。人的手指运动非常复杂，很多操作不仅涉及动作轨迹，还涉及触觉、力反馈以及不同关节之间的协调。要把这些信息准确采集下来，并不容易。此外，复杂操作的数据采集效率也是问题。例如，电力系统带电作业场景需要机器人在特定位置完成精细操作。这样的任务并不是通过简单遥操作就能获得大量高质量数据，实际采集过程也存在很多困难。
Current embodied intelligence's main technical route remains end-to-end data-driven methods. Data collection is necessary, but which collection method is most effective is still hard to judge. Our laboratory also uses different methods including motion capture and VR devices. Different collection methods have their own advantages and limitations. For example, some methods are effective for collecting arm and body joint movements but may not be ideal for fine finger movements. Human finger movement is very complex; many operations involve not only motion trajectories but also touch, force feedback, and coordination between different joints. Accurately collecting this information is not easy. Additionally, data collection efficiency for complex operations is also a problem. For example, power system live-line work scenarios require robots to complete fine operations at specific positions. Such tasks cannot obtain large amounts of high-quality data through simple teleoperation; actual collection also faces many difficulties.
这实际上涉及机器人学习方式的问题。人学习一件事情，并不需要在每一种情况下都重复训练。比如，今天桌子上放着一个杯子，你学会了如何拿起它。明天换成另一个杯子，或者桌子上多了一本书，你不需要再从头学习一遍。但如果我们的机器人依赖大量数据训练，那么环境发生一点变化，就可能需要重新采集数据、重新训练，这显然不是理想的方式。所以，我们需要更多借鉴人脑的工作机制，比如在线校正，以及基于内部模型的预测等机制。机器人在执行任务的过程中，应当能够根据实际情况不断调整自己的行为，而不是仅仅按照事先训练的方式完成动作。只有将这些机制引入具身智能系统，才有可能减少对大量训练数据的依赖，并且更有效地完成复杂任务。
This actually involves the question of robot learning methods. Humans learning something do not need to repeatedly train in every situation. For example, today there is a cup on the table and you learn how to pick it up. Tomorrow it is a different cup, or there is an extra book on the table—you do not need to learn from scratch again. But if our robots rely on massive data training, then a slight environment change may require re-collecting data and retraining—clearly not ideal. So we need to borrow more from how the human brain works, such as online correction and internal model-based prediction mechanisms. Robots executing tasks should be able to continuously adjust their behavior based on actual conditions rather than merely completing actions according to pre-trained patterns. Only by introducing these mechanisms into embodied intelligence systems can we reduce dependence on massive training data and more effectively complete complex tasks.
中国很大的一个优势，就是制造业体系比较完整，有非常丰富的工业应用场景。同时，政府也比较重视具身智能产业发展。这些条件对于具身智能发展非常重要。不过，我们现在很难期待机器人像人一样，能够完成各种各样的任务。这样的目标想要在短期内达成不太现实。因此，可以先从一些垂直领域入手。但我们真正希望实现的是，机器人具备一定的自主学习能力，在面对不同型号产品、不同工作条件的时候，不需要每一次都从头进行训练。如果能够实现这一点，对于机器人在实际生产中的应用就会很有帮助。结合具体行业和产品开展训练与研发，是值得重视的方向。
China's big advantage is a relatively complete manufacturing system with very rich industrial application scenarios. Meanwhile, the government also attaches importance to embodied intelligence industry development. These conditions are very important for embodied intelligence development. However, we can hardly expect robots to complete all kinds of tasks like humans now. Such a goal is unrealistic to achieve in the short term. Therefore, we can start from some vertical fields. But what we truly hope to achieve is robots possessing certain autonomous learning capability, so that when facing different product models and different working conditions, they do not need to retrain from scratch every time. If this can be achieved, it would be very helpful for robots' practical production application. Conducting training and R&D combined with specific industries and products is a direction worth emphasizing.
如果从未来5至10年的产业竞争来看，具身智能企业的核心竞争壁垒会建立在什么方面。金耀初认为，这些方面都需要。首先，硬件依然非常重要。从成本和部分功能表现来看，中国机器人硬件在国际上已经具有一定竞争力。但在实际使用过程中发现，硬件本身仍然存在很多局限性。例如稳定性和可靠性不足，感知能力也有待提升。特别是在灵巧手方面，机器人的手指控制、传感能力仍然存在问题。有些机器人虽然能够完成动作，但是手指的驱动和控制方式还比较有限，传感信号的质量也有待提升。这些问题都会影响机器人的操作能力，甚至影响数据采集和后续模型训练的效果。所以，不能只关注人工智能算法，还需要重视硬件与算法之间的匹配。如果硬件不够好，即使算法再先进，也很难发挥出应有的能力。
Looking at industrial competition over the next 5 to 10 years, where will embodied intelligence enterprises' core competitive barriers be built. Jin believes all these aspects are needed. First, hardware remains very important. From cost and some functional performance perspectives, Chinese robot hardware already has certain international competitiveness. But in actual use, hardware itself still has many limitations. For example, insufficient stability and reliability, and perception capability needs improvement. Especially in dexterous hands, robot finger control and sensing capability still have problems. Some robots can complete actions, but finger driving and control methods are still relatively limited, and sensing signal quality needs improvement. These problems affect robot manipulation capability and even data collection and subsequent model training. So we cannot focus only on AI algorithms; we must also value hardware-algorithm matching. If hardware is not good enough, even the most advanced algorithms struggle to deliver their due capability.
还需要提升数据采集、感知以及控制等技术能力。再进一步，就是类脑思想和闭环机制，让机器人能够在执行任务时根据环境变化自主调整。只有硬件、传感、控制和人工智能技术协同发展，才有可能让具身智能可靠地应用于更多复杂场景。关于是否需要放缓AI发展的速度，金耀初认为，简单地踩刹车没有什么用。更重要的问题是我们应该怎样开发更加安全、可信、可靠的人工智能系统。人工智能技术需要继续发展，其内生安全性也需要同步提升。特别是具身智能，它与传统人工智能有所不同。机器人不仅能够作出决策，还能够在物理世界中执行动作，潜在的安全风险可能更加直接。因此，在开发人工智能技术的时候，就应该考虑如何提高其安全性和可靠性，而不是等到技术成熟之后，再去补充安全方面的能力。
We also need to improve data collection, perception, and control technical capabilities. Further, brain-inspired thinking and closed-loop mechanisms let robots autonomously adjust based on environmental changes during task execution. Only with coordinated development of hardware, sensing, control, and AI technology can embodied intelligence be reliably applied in more complex scenarios. Regarding whether AI development speed should be slowed, Jin believes simply "stepping on the brakes" is useless. The more important question is how we should develop safer, more trustworthy, and more reliable AI systems. AI technology needs continued development, and its intrinsic safety needs simultaneous improvement. Especially embodied intelligence differs from traditional AI. Robots can not only make decisions but also execute actions in the physical world, with potentially more direct safety risks. Therefore, when developing AI technology, we should consider how to improve its safety and reliability rather than waiting until technology matures to add safety capabilities.
