Dyna Robotics 首席科学家 Jason Ma：百万小时世界-动作模型与具身 AI

Jason Ma

若要学一个新任务，我们会做一点点机器人数据采集，然后模型就能用。多少算「一点点」？大概几小时；有些任务只要几十分钟——切菜那个大约 45 分钟，开冰箱也类似，几乎是工程师入职任务级别的数据量。

If we want to learn a new task, we basically do a little bit of robot data collection, then the model… What is a little bit? Like a couple hours. In some of the tasks, just tens of minutes. The chopping one was 45 minutes. The open-the-fridge one—that’s like an onboarding task for one of your engineers.

Jason Ma

我们的机器人今天在美国很多不同地点为客户折数千张餐巾。我想强调：很多机器人公司在谈部署、谈试点；真正难的不是把机器人放进一个点，而是客户验证它创造了足够价值、愿意在很多门店铺开。据我所知，做类似通用机器人的公司里，很少有人真正跨过这一步——鼎泰丰已经和我们走到这里。

Our robots today fold thousands of napkins for many different locations in the United States. A lot of robotics companies talk about deployments and pilots, but what’s really hard is not putting your robot in one location. What’s really hard is that the customers validate that your robot creates enough value that they want to deploy across many of their sites. To the best of my knowledge that hasn’t really happened for other companies doing similar general-purpose robots. That’s what’s happened for us with Din Tai Fung.

swyx

我们在现场采访 Dyna Robotics 联合创始人、首席科学家 Jason Ma。镜头外看到了取饮料、用刀切菜。底层是世界-动作模型，那是什么？

We’re here on site with Jason Ma, co-founder, chief scientist of Dyna Robotics. We got to see drinks being pulled, robots with knives. Under the hood you have a world-action model. What is that?

Jason Ma

上个月我们发布了世界-动作模型 Dyna-2，据我们所知是首个在超过 100 万小时数据上训练的世界-动作模型。世界模型能预测未来视频、世界未来状态；此外还可以用这种未来生成，产出机器人能在物理世界执行的动作。人类行动时脑子里也在隐式想「世界会如何随我的动作演化」；有了这种架构，模型更灵巧、更稳健，也更能泛化到从未见过的场景。

Last month we released our world-action model called Dyna-2. It’s the first world-action model trained on more than one million hours of data. A world-action model predicts the future video, the future state of the world — hence world model — but in addition you can use that future generation to produce actions robots can execute. In our brain we are implicitly thinking about how the world will evolve based on our actions. By doing so the model is much more dextrous, much more robust, and much more capable of generalizing to environments and scenarios it’s never seen before.

Jason Ma

建机器人模型主要看两件事：数据，以及如何让模型从数据里学到尽量多。Dyna-2 底层是扩散 Transformer——序列建模上像你熟悉的 Transformer，但在视觉域，生成图像、视频、以及从图像生成动作时，人们发现扩散类架构特别有效。扩散从随机噪声出发，经迭代算出高维、细粒度的结果；图像要锐利边缘与对比度，机器人虽不生成图像，但生成动作同样需要细粒度。

There are essentially a couple things: data, and how we make a model learn as much as possible from that data. Under the hood Dyna-2 is a diffusion transformer. In visual domains — generating images, videos, and in this case robots looking at images to generate actions — people have found the diffusion kind of architecture to work really well. A diffusion model takes random noise and through iterative computation goes to something high-dimensional with fine-grain features. For robotics, even though you’re not generating an image, you’re generating actions that need that kind of detail.

Jason Ma

今天机器人里最重要的是数据，因为不像语言模型，这里没有互联网。每个做模型、做硬件的公司都必须想数据。你可以自己采很多——我们办公室到处是机器人——但单个公司能采到的量，对机器人基础模型远远不够，就像想用 Reddit 或 Facebook 一家站点拼出整个互联网。所以还要买数据、用已有数据。

Data is the most important thing in robotics today because unlike language models, there is no internet. Every company building models and hardware has to think about data. You can collect a lot yourself — we have robots everywhere in the office — but the amount any individual company can collect is nowhere near sufficient for a robotic foundation model. It’s like trying to build an internet by just having Reddit or Facebook. There has to be data acquisition: you buy data, you look at data that’s already out there.

Jason Ma

机器人里的「数据质量」是非平稳的：模型什么都不会时，任何数据都有用，可以用大量视频；模型变强后，什么叫高质量会变。

Data quality in robotics is a non-stationary problem: what is high-quality data for your model changes over time. When your model doesn’t know anything, any data is useful — you can use a lot of video data.

Jason Ma

为机器人生成合成环境、合成数据、仿真数据，会是下一扩展前沿。自我中心视频、机器人数据在扩展，合成数据会是新维度。我们在用前沿模型为我们关心的场景生成合成环境；出乎意料的是，技术已经到了值得认真深挖的程度。

Generating synthetic environments, synthetic data, simulation data for robotics is going to be one of the next frontiers. Egocentric data and robot data are scaling, but synthetic data will be a new dimension of scaling. We’ve been exploring using frontier models to generate synthetic environments for the data and environments we care about, and surprisingly it’s at a point where this technology is very much worth going really deep into.

Jason Ma

我博士主要做基于仿真的机器人训练。那时有 Eureka 论文：用当时的 GPT-4 自生成环境与奖励做强化学习，最后有机器人转笔的演示。但当时模型视觉弱、没有内建 3D 理解，限制很大；现在不一样了。

In my PhD I primarily focused on simulation-based training for robotics. We had this paper called Eureka — the first paper taking a large language model, GPT-4 at the time, to self-generate environments and reward functions for reinforcement learning. We had a nice demo of the robot spinning pencils. But that was still very limited because the model didn’t have good vision or an innate 3D understanding.

Jason Ma

机器人和语言模型很不一样：编码智能体、聊天可以有几秒延迟，机器人不行。你必须在物理世界实时出动作，所以约束更像端侧 / 混合部署，而不是「把参数和数据无量堆上去」。百万小时很好，但扩展定律要受实时推理约束。

What’s really different about robots versus language models is that for coding agents and language models it’s okay if there’s latency — ChatGPT can take a couple of seconds. For robots you need real-time actions in the physical world. That constrains how you scale parameters and data; a million hours is great, but you operate under on-device and hybrid inference constraints.

Jason Ma

什么自研、什么外购：若一个组件的输入输出非常清晰、已有成熟方案，就不必自建——比如机器人上的激光雷达，自动驾驶用了几十年，我们不必从零做雷达；部分相机也类似。但对系统最终性能极其关键的，我们希望自己掌握。执行器就是一例：执行器质量与设计决定你能不能在仿真里把机器人建模好；建模不好，后面一切都难。比例和尺寸我们对照真人量过，因为我们在人类数据上预训练，迁移更好。硬件是承载智能的，必须为智能服务。

If the input/output of a component is very clear and there’s already an extremely good mature solution, there’s no point building it in-house — lidars help detect obstacles and navigate; that’s mature technology from self-driving, so we don’t build lidar from scratch. Same with some cameras. But for anything extremely crucial to end-to-end performance, our vision is to own it. Actuators: the quality and how you design the actuator really matters for how well you can model the robot. Proportions we measured against real people because we’re pre-training on human data — best transfer. Eventually it’s hardware that carries intelligence.

Jason Ma

我们把平台看成半人形：上半身是人形——肩膀、手臂，腰部还有四个自由度，能做人类能做的上半身动作；下半身是轮子。为什么用轮？实用与性能的权衡：腿式能耗高得多、更不稳，安全极难保证；今天很多腿式机器人并不能全程保证直立平衡，跳舞视频里也很容易倒。洗衣演示里要探进机器，多出来的腰部自由度非常关键——日常里很多你习以为常的肌肉与自由度，一旦「落枕」或缺了某轴，做平常任务都会别扭。

We think of this as generally a semi-humanoid: the upper body is like a humanoid — shoulders, arms, and four degrees of freedom on the waist so you can do all the upper-body motion a human can. It’s semi-humanoid because the lower body is wheels. Why wheels? Optimal practicality versus performance: with legs, energy consumption is a lot higher, you’ll be a lot more unstable, and safety is extremely hard to guarantee. For the laundry demo you have to lean into machines — that’s another degree of freedom. When your neck is stiff you realize how many degrees of freedom you take for granted.

Jason Ma

学新任务仍然是少量机器人数据加已有世界-动作模型的泛化；评估上我们用很多「好玩但能压测能力」的任务——例如按语言指令抓特定饮料，就是在测语言跟随与意图理解。

Learning a new task is still a little robot data plus the generalization of the world-action model. For evals we create tasks that stress-test capabilities — grabbing a specified drink tests language following: if the model follows intent and instruction, it should be able to do it.

Jason Ma

我认为通向更通用能力的最大瓶颈，仍然是硬件和数据。模型可以很强，但若身体够不着、传感器不对、或数据覆盖不到形变物体的无穷构型，部署就会卡在物理现实上。

I do think the biggest bottleneck to having a much more general thing is still hardware and data.

Jason Ma

我们和酒店业的 Best Western 等合作，帮他们叠大量毛巾——酒店、健身房、健身中心都需要。网站上有部署博文和客户合集视频。鼎泰丰这类偏高档餐厅桌上有餐巾，我们的机器人在美国多个地点每天折数千张；质检极严，比如第二道折缝要在袋口下方几厘米，好插筷子。可变形物体有无穷构型，专用机很难；用通用模型再适配到场景，反而更快到生产质量。鼎泰丰规模化并签了更大合同后，不只折餐巾，还希望机器人做切菜等更多事。

We have customers like Best Western in hotels — we fold a lot of laundry, towels; every hotel and gym needs it. At Din Tai Fung, if you go to a somewhat upscale restaurant there are napkins on the table. Our robots fold thousands of napkins per day across many locations in the United States. The quality standard is extremely high — the second forward seam only a couple centimeters below the pocket so you can stick chopsticks in. Manipulating deformable objects — napkins, towels — there are infinitely many configurations, which is really hard for a specialized robot. By developing a general-purpose model and adapting it, you can get to production quality faster. Now that Din Tai Fung has scaled and signed a much bigger contract, they want robots not only folding napkins but cutting and more.

Jason Ma

机器人也会像机器学习解决过的其他领域一样扩展——尽量多要正确数据。但「什么是正确数据」比别的领域难得多：语言是文本，游戏是游戏数据；机器人里你选什么传感器，就决定你有什么数据。传感器和硬件选对了再扩展，模型性能会有可预期回报。苦涩教训的推论是：别堆「先画框再感知」这类很强的专用先验，保持算法更简单，扩展正确数据，更好的泛化与行为会涌现。扩展终将奏效，但机器人没有互联网，必须先把硬件、传感器做对再扩展，难得多。

Robotics is going to scale just like any other domain machine learning has solved: have as much of the right data as possible. But what the right data is is much harder than in other domains. In language it’s text; in video games, game data; in robotics, what sensors you choose determine what data you have. If you have the right sensors and scale that data, you get predictable return on model performance. The bitter lesson corollary: there’s a lot of research coming up with very specific priors — draw boxes so the robot sees objects better — but keep the algorithm simpler and scale the right kind of data, then better generalization emerges. Scaling will work, but how to scale is much harder because there is no internet: you have to get hardware and sensors right, then scale.

Jason Ma

若目标是尽量多场景可适配的通用机器人，长远看人形外形可能最好，因为世界是为人造的；但必须对当下硬件务实，所以我们现在走轮式半人形。

If you want a general-purpose robot that can adapt to as many use cases as possible, I think human form factor will eventually be the best one, because that’s what we’ve been doing for ages. But you have to be practical with what we have now.

Jason Ma

我们想要一层通用的机器人基础模型「外壳」（harness）；今天里面是我们自己的模型。物理世界里的长任务有大量高层决策，不只是输出连续动作轨迹。洗衣流程里：洗衣机在转就该去干别的；洗完了要立刻把毛巾拿去烘干，否则整条流水线卡住。理论上可以只靠底层动作模型学这些，但把高层认知放进外壳会有效得多。最终我们也会有自己的 SDK。

We want a general harness layer for a robotics foundation model — today that’s our own model. For any real long task in the physical world there’s a lot of high-level decision-making, not just continuous action trajectories. For laundry: the washer is running, so go do something else; when it’s done, take towels to the dryer immediately or that becomes a bottleneck. In theory you could try to learn that with a low-level action model, but you need that high-level cognition. Eventually we will have our own SDK.

Jason Ma

语言模型与编码工具变强，深刻影响我们。我们是极度 AI 原生的公司，编码智能体是工作的基本盘；新模型出来我们往往第一周就全员黑客松。GPT-6 出来那周的黑客松里，工程师做的东西很惊艳，很多已经进主线和生产。GPT-6 让我们兴奋的一点是它构建虚拟环境、变成仿真的能力——和二十分钟前说的合成数据是同一方向。

It affects us very deeply. We’re an extremely AI-native company; coding agents are a fundamental part of our work. Whenever there’s a new model, we’re the first to test it. When GPT-6 came out we had a company-wide hackathon the same week, and what our engineers built was incredible — many features are already in main, in production. What’s exciting about GPT-6 is the ability to construct virtual environments that turn into simulation — the synthetic-data direction we talked about.

Jason Ma

机器人是今天最兴奋的领域之一，很多公司在做前沿工作。路线仍很分歧：我们相信硬件要自研、要吃下整条链，并认为扩展人类数据极其重要；也有公司押注只扩展机器人数据，或扩展某种中间采集装置。现在还早，看谁跑出来。我们自我感觉良好的一点是：公司使命不只是「做一个模型」，而是真把机器人部署出去——部署越多，越早撞上难题。

Robotics is the most exciting space today, and there are many amazing companies doing frontier work. There are still many approaches: we believe hardware needs to be built in-house and we need to own the whole chain; we believe scaling human data is extremely important. Other companies bet on just scaling robot data, or intermediate data-collection devices. It’s still early. We feel pretty good because culture- and mission-wise we’re not here just to build a model — we’re here to get robots deployed. We deploy as much as possible, work on the hard problems, and identify issues earlier.

Jason Ma

我在日本和中国待过，从小下围棋。AlphaGo 出来时我高三，整个人被震撼——以前人们说围棋搜索空间大于宇宙原子数，AI 永远解不了，结果强化学习直接凿开。我因此对能自我改进的智能体着了迷，到宾夕法尼亚大学读博就想做强化学习；联导里有一位偏机器人、在伯克利机器人实验室做过博后，我便进了 GRASP，做面向机器人的 RL，以及如何把 RL 与人类数据、严谨理论连起来。

As a kid I spent time in Japan and China. I grew up playing Go. When AlphaGo came out I was a senior in high school and I was mind-blown — people said Go was a game AI would never solve, search space larger than atoms in the universe, then RL cracked it. That’s how I got into AI: systems and agents that can self-improve. At Penn as a PhD student I really wanted to work on RL. One advisor was more in robotics — postdoc at the Berkeley robotics lab — so I got into robots at GRASP, connecting RL with human data and doing RL theory rigorously.

Jason Ma

后来去 NVIDIA，那时 GPT-4 刚出，和 Jim Fan 等人想做灵巧转笔：起初想从人类视频学，但发现没有能转笔的灵巧手，算法好不好都无法评。于是先上仿真，立刻发现手写转笔奖励几乎不可能——这才变成 Eureka：让会写代码的模型来写环境与奖励。在我看来那是仿真里第一批递归自我改进的机器人智能体之一，让我看到语言模型/编码模型加速机器人的潜力。再往后去了 Google DeepMind，大约 2024 年底出来创办 Dyna。

Then I went to NVIDIA when GPT-4 came out. Working with Jim Fan and others, we wanted dexterous pen spinning. The initial idea was learning from human videos, but we didn’t have a dextrous hand that could spin a pencil at all, so we couldn’t evaluate the algorithm. In simulation I immediately saw the issue: coming up with the reward function for pen spinning by hand is impossible. That became Eureka — what if a model that can write code does that. In my opinion we had one of the first recursive self-improving robotic agents in simulation. Then I went to Google DeepMind, and later founded Dyna.

Jason Ma

我们把自己想成最终的「研究+产品」公司，有点像 OpenAI：做前沿研究，但最终都要进到产品里。我们不是开放式地今天对什么好奇就做什么，而是产品驱动——要给新客户部署机器人/平台，必须解决泛化、全身控制、快速适应、语言跟随、灵巧。问题清楚后，能干的工程师会自底向上发明压测这些能力的任务；抓饮料就是测语言跟随。大家长期热爱机器人，文化让这些事自然发生。

We think of ourselves as eventually a research-plus-product company, kind of like OpenAI — frontier research that makes its way into the product. We’re not here to do open-ended whatever interests us; we’re product-driven. We want to deploy robots to a new customer: we have to solve generalization, whole-body control, adapting quickly, following language, dexterity. Once those are clear, engineers create tasks that stress-test those capabilities. The grab-a-drink task tests language following. Everyone here is super passionate about robots; we create a culture where it just naturally happens.

Jason Ma

公司最近长很快，我没有精确比例，硬件侧大概 30%–40%，因为硬件更吃人；模型大约 10 人左右，还有软件/机器人工程师和很强的业务侧。做机器人公司最有趣的是：你能碰到太阳底下一切可能有用的东西，暴露面极大。

We recently grew so much, so I don’t have the perfect number, but on the hardware side probably like 30% to 40%, because hardware needs a lot more people. Around 10 people working on models, plus software engineers, robotics engineers, and very talented people on the business side. The most fun thing about building a robotics company is you see everything under the sun that could possibly be helpful.