字号·· | 护眼
经济学人

何小鹏:为人形机器人做好准备Get ready for humanoid robots, says He Xiaopeng

点「原文对照」整页切到原文,或双击某段只看那段的原文。

人形机器人行业有一个奇特的习惯。它喜欢一次只造一个身体部位来构建未来。一些公司专注于运动控制,而另一些则致力于灵巧手或通用大脑。几乎所有人都在谈论数据,往往假设数据越多就是灵丹妙药。然而,如果数据量本身就是答案,那么坐拥海量真实驾驶数据的车企早就该交付完全自动驾驶汽车了。但它们并没有。

The humanoid-robot industry has a peculiar habit. It likes to build the future one body part at a time. Some companies focus on motion control, while others devote their attention to dexterous hands or general-purpose brains. Almost everyone talks about data, often assuming that more is a panacea. Yet if data volume alone were the answer, then carmakers, sitting on oceans of real-world driving data, would have already delivered fully autonomous vehicles. They have not.

如今大多数人形机器人不过是远程操控的傀儡或昙花一现的演示机型。它们的存在是为了取悦消费者或给投资者留下深刻印象。从演示走向可靠部署是一个巨大的飞跃。要让这些机器从会议舞台走进工厂、办公室甚至人们的家庭,我们需要正确的架构。它立足于三大支柱:硬件、软件和数据。

Most humanoid robots today are not much more than tele-operated puppets or short-lived demonstration models. They exist to entertain consumers or impress investors. To advance from demonstration to dependable deployment is a huge leap. For these machines to go from conference stages to factories, offices and even into people’s homes, we need the correct architecture. It rests on three pillars: hardware, software and data.

机器人必须以软件思维设计,但硬件必须先行。在汽车行业,硬件先于软件,两者的界限相对清晰。人形机器人则不同。它们的硬件与智能必须从一开始就协同构想。这容易描述,却极难工程化落地。行业发展太快,导致硬件和软件都难以大规模获取。

Robots must be designed with a software mindset, but hardware must come first. In the car industry, hardware precedes software, and the boundary between the two is comparatively clear. Humanoid robots are different. Their hardware and intelligence must be conceived together, right from the start. This is easy to describe, but extraordinarily difficult to engineer. The industry is moving so fast that both hardware and software are difficult to source at scale.

相反,成功将要求两者都在内部自主研发,采取全栈式的研究与创新方法。一辆典型的电动汽车有一到两个驱动电机;小鹏的人形机器人却有76个,每个关节执行器对应一个。实现安全与可靠,要求的严谨程度远超造车。

Instead, success will require both to be developed in-house, with a holistic approach to research and innovation. A typical electric car has one or two drive motors; Xpeng’s humanoid robot has 76, one for each joint actuator. Achieving safety and reliability demands a level of rigour far beyond that needed in carmaking.

在软件层面,许多企业希望成为机器人领域的“安卓”,打造一个单一的 AI 模型来统筹从视觉到运动的一切。当前的模型在智能和泛化能力上表现出色。但正如 2018 年的自动驾驶汽车,它们在演示中看起来不错,却远未达到在成本、安全、精度和可靠性之间取得平衡的程度。

When it comes to the software, many firms hope to be the “Android” of robotics, creating a single AI model to orchestrate everything from vision to movement. Today’s models shine in intelligence and generalisation. But like self-driving cars in 2018, they look good in demonstrations, yet are far from balancing cost, safety, precision and reliability.

与其追求一个万能的大脑,人形机器人将需要几种不同速度的思维:用于运动控制和瞬间反射的超快模型;用于运动和实时交互的快模型;用于推理和记忆的慢模型;以及专门负责安全的独立模型。这正是我们在小鹏采取的路径。人类的运作方式大致相同。我们不会在把手从烫手的表面上缩回之前先写一篇文章。反思有其位置,反射同样有其位置。

Rather than one all-purpose brain, humanoid robots will need several speeds of thought: ultra-fast models for motor control and split-second reflexes; fast models for motion and real-time interaction; slow models for reasoning and memory; separate models dedicated to safety. This is the approach we are taking at Xpeng. Humans work in much the same way. We do not write an essay before pulling our hand away from a hot surface. Reflection has its place, but so do reflexes.

潜在收益最大的场景,却是最不宽容的工程环境 就数据而言,训练一个典型的大模型大约需要 200 TB(太字节)数据。训练一个用于控制自动驾驶汽车的“视觉-语言-动作”模型需要的数据量是其 1000 倍,即约 200 PB(拍字节)。而训练人形机器人将需要更多的数据。然而,仅凭规模告诉我们的信息非常有限。

The setting with the greatest potential benefit is the least forgiving engineering environment As for data, training a typical large model requires roughly 200 terabytes of data. Training a “vision-language-action” model to control an autonomous car requires 1,000 times more, or around 200 petabytes. And training a humanoid robot will require even more than that. Yet scale alone tells us very little.

数据需要多样性,而不仅仅是体量。目标应当是构建高质量的数据闭环和基础设施,形成数据与能力自我强化的飞轮效应。在小鹏,我们认为用于训练汽车的系统可以直接应用于机器人。本体可能不同,但工程逻辑具有很好的迁移性。

Data needs variety, not merely volume. The objective should be to build high-quality data loops and infrastructure that create a self-reinforcing flywheel of data and capability. At Xpeng, we think the systems we built to train cars can be applied directly to robots. The body may be different, but the engineering logic travels well.

这三大支柱仅是基础。对于通用人形机器人而言,还必须在另外四个领域取得进展。机器人必须像儿童一样,学会如何安全移动。它们必须学会使用工具,并在各类环境中导航。它们必须能够学习并应用专业知识与技能。此外,它们还需要具备社交与情感智能,以便根据情境做出恰当行动。

These three pillars are just the foundation. For a general-purpose humanoid robot, there will also have to be progress in four other areas. Robots must learn, just as children do, how to move around safely. They must learn to handle tools and navigate a range of environments. They must be able to learn and apply specialist knowledge and skills. And they will need social and emotional intelligence to act correctly depending on the situation.

为实现这一切,人形机器人应分阶段投入应用。商业用途——如为购物者或游客提供向导服务——是合乎逻辑的切入点。此类任务往往重复性强、定义明确,且在受控环境中执行。工业场景则引入了更高的复杂性。家庭环境是最具挑战性的场所:私密、非结构化,且充满美妙的不可预测性。机器人领域的悖论在于:最能造福人类的场景,却是最不宽容的工程环境。

To achieve all this, humanoids should be deployed in stages. Commercial uses, such as providing guidance to shoppers or tourists, offer a logical starting-point. Tasks tend to be repetitive, clearly defined and performed in controlled surroundings. Industrial settings introduce greater complexity. The home is the most challenging environment of all: intimate, unstructured and gloriously unpredictable. The paradox of robotics is that the setting with the greatest potential human benefit is the least forgiving engineering environment.

挑战艰巨。但我坚信,2027年将标志着先进通用人形机器人迈入大规模量产的开端。在小鹏,我们正致力于攻克实现这些目标所需的核心能力。对于机器人产业而言,我们的雄心不应止步于制造强大的机器,而必须确保它们丰富人类生活、拓展人类身体能力——让人们获得更多自由,专注于唯有人类才能做的事情。■

The challenges are formidable. But I believe that 2027 will mark the beginning of mass production for advanced general-purpose humanoid robots. At Xpeng we are currently pursuing the capabilities needed to meet these goals. For the robotics industry, our ambition must not be simply to build capable machines. It must be to ensure they enrich human life and extend our physical capabilities—giving people more freedom to focus on what only humans can do. ■

阅读媒体原文 ↗