写给大模型与 AI 从业者:通用模型正在抬高机器人研发的起点,而最值得亲手搭建的,是让能力在真实世界里持续生长的闭环。For people who build large models: general models are raising the starting line for robotics, and the thing most worth building with your own hands is the loop that lets that capability keep growing in the real world.

Read in English →阅读中文版 →

2026 年 11 月 13 日,《财富》报道了一则乍看矛盾的消息:OpenAI 解散了独立的机器人团队。原因不是放弃机器人,而是它的通用模型已经开始直接驱动机器人。

机器人研究员并入首席研究官 Mark Chen 领导的核心研究团队,硬件与应用机器人团队转入 Greg Brockman 负责的应用部门。OpenAI 表示,这样做是为了加快、而不是放弃机器人方向;据报道,研究人员把 GPT-6 Astra 接到机器人上时,看到了出人意料的表现。卡内基梅隆大学的 Guanya Shi 说得更直白:在 OpenAI 眼里,机器人已经“不再是一个特殊问题”。

这件事让我更加确信一个判断:现在,可能是迄今为止大模型与 AI 从业者进入具身智能最好的时候。

这不是说机器人变容易了。接触、动力学和实时控制依然很难,依然需要深厚的专业积累。变化发生在入口:通用模型开始把理解、推理与行动连成一体,让我们看到了直接以大模型为能力基础构建机器人的曙光。

On November 13, 2026, Fortune reported news that sounds like a contradiction: OpenAI had disbanded its stand-alone robotics team. Not because it was giving up on robots, but because its general models had started running them.

Its robotics researchers were folded into the core research team under chief research officer Mark Chen, while hardware and applied robotics moved into the Applied group under president Greg Brockman. OpenAI framed the move as a way to accelerate robotics, not retreat from it, and according to the report, its researchers had seen surprising performance when they connected GPT-6 Astra to robots. Carnegie Mellon’s Guanya Shi put it more bluntly. To OpenAI, robotics is “no longer a special problem.”

The news sharpened a conviction I have been forming for some time: this may be the best moment yet for people who work on large models and AI to step into embodied intelligence.

That is not because robotics has become easy. Contact, dynamics and real-time control are as hard as ever, and they still reward deep expertise. What has changed is the way in. General models are beginning to join understanding, reasoning and action into one system, and we can now see the first light of robots built directly on a foundation model’s capabilities.

起点 · 为什么是现在The opening · Why now

为什么这个窗口正在打开Why the window is opening

我的理解是,具身正在成为通用大模型的一个重要应用域,一个新的 subdomain。编程已经走过这条路:专门的代码模型曾各领风骚,如今通用模型的每一次进步,几乎都会直接变成更强的编程能力。如果机器人沿着同样的路线发展,它也会持续受益于通用模型在理解、推理和学习上的进步。模型每进步一次,机器人研发的起点就被抬高一次。

信号也不只来自一家。Google DeepMind 的 Gemini Robotics 把物理动作变成了模型的一种输出;Anthropic 报告,Fable 5.2 在仿真机器人基准 RoboDojo 上超过了同一测试框架下的其他前沿模型,也超过了多数专为机器人打造的系统;《财富》还援引了一项 NVIDIA 的研究:在一个基于仿真的操作测试中,GPT-6 Astra 驱动的智能体成功率为 74.8%,而榜单上最好的机器人专用模型是 51.0%。这些都是仿真或厂商自报的结果,不能等同于真实机器人的成功率,但方向相当一致:在这些测试里,通用模型已经开始追上、甚至超过专用系统。

这正是窗口期的意义。过去,进入一个机器人任务,往往要先在感知、规划和控制上积累大量专门知识,再把一个个模块手工连接起来。通用模型让其中相当一部分知识与能力可以复用,不必再为每个任务从头设计一条专用流程。对熟悉模型训练、多模态、智能体和推理系统的团队来说,已有的积累因此更容易转化为进入具身的起点。

所以,我说的“最好的时候”,是指通用能力的增长正在打开新的入口。机器人的接触、动力学和实时控制仍然需要专业知识,但这些知识可以逐步沉淀到模型、技能与系统接口中,让更多 AI 团队站在已有的基础上开展工作。

入口变宽之后,稀缺的东西也变了。MIT 的 Russ Tedrake 在同一篇报道中提醒:已经在工厂里干活的机器人,会持续产生只属于那个现场的数据,而这些数据会随着基础模型变强而越来越值钱。我深以为然。

My reading is that embodiment is becoming a major application domain of general foundation models, a new subdomain. Coding has already walked this path. Specialized code models once had their moment; today nearly every advance in general models turns directly into better coding. If robotics follows the same route, it will keep inheriting progress in understanding, reasoning and learning. Every time the base model improves, the starting line for robot development moves forward.

The signal is not coming from one lab. Google DeepMind’s Gemini Robotics made physical action an output of the model. Anthropic reports that Fable 5.2 outperformed the other frontier models run on the same harness on the simulated RoboDojo benchmark, as well as most systems built specifically to control robots. Fortune also cites an NVIDIA study in which a GPT-6 Astra–powered agent succeeded 74.8% of the time on a simulation-based manipulation test, against 51.0% for the best specialized robot model on that leaderboard. These are simulated or self-reported results, not real-robot success rates. But they point the same way: in these tests, general models are catching up with, and sometimes overtaking, specialized systems.

That is what makes this a window. Entering a robotics task used to mean first accumulating a stack of specialized knowledge in perception, planning and control, then wiring the modules together by hand. General models let much of that knowledge be reused, so you no longer need to design a bespoke pipeline for every task. For teams fluent in model training, multimodality, agents and inference systems, what they already know becomes a far shorter path into embodiment.

So when I say “the best moment,” I mean that the growth of general capability is opening new doors. Contact, dynamics and real-time control still demand expertise, but that expertise can be steadily distilled into models, skills and system interfaces, letting more AI teams build on what already exists.

When the door widens, what is scarce changes too. In the same Fortune piece, MIT’s Russ Tedrake points out that robots already working on factory floors keep generating data specific to their sites, data that grows more valuable as base models improve. I could not agree more.

模型可以借用,闭环必须自建。You can borrow the model. You have to build the loop.

通用模型会一代代更替,但你在真实场景中搭起来的数据、评价与部署体系,会把每一代新模型的能力放大。接下来,我沿着三个层层外扩的闭环展开:

General models will turn over generation after generation. The data, evaluation and deployment you build in a real setting will amplify every one of them. The rest of this essay follows three loops, each wrapping around the last:

闭环一 · 模型Loop 1 · The model

通用能力怎样变成行动能力How general capability becomes the ability to act

共享的能力,抬高每个新任务的起点Shared capabilities raise the floor for every new task

我看好通用模型走向具身,是因为机器人面对的任务里,有大量能力是可以共享的:理解人的意图,辨认环境中的关系,拆解目标,选择工具,再根据结果调整计划。如果这些能力能随通用模型一起进步,每进入一个新场景,研发就能从更高的起点出发。

这也改变了我对机器人模型研发的理解。值得追求的,不是一个只会完成既定任务的专用模型,而是一个能够继续学习、适应新任务的能力基础。模型对世界理解得越充分,特定场景需要补充的经验就越会集中到真正困难的地方:这副身体能做什么,动作会怎样改变环境,出现偏差时如何恢复。

I am optimistic about general models moving into embodiment because so much of what robots need is shared: understanding what a person wants, recognizing relationships in a scene, breaking a goal into steps, choosing tools, and revising the plan based on what happened. If those capabilities improve along with general models, every new setting starts from a higher floor.

This has changed how I think about building robot models. What is worth pursuing is not a specialist that completes a fixed set of tasks, but a foundation of capability that keeps learning and adapting. The better a model understands the world, the more the experience it still needs concentrates on what is genuinely hard: what this particular body can do, how its actions change the environment, and how to recover when things drift off course.

让行动的后果进入学习Letting consequences into learning

能力可以共享,行动的后果却只能在具体环境里学会。语言世界是宽容的:一句话说错了,下一轮还能改口。物理世界没有撤销键:杯子打翻了,就是打翻了。一次动作会改变下一次观察,前面累积的误差会拖累后面的计划,执行的时机本身就决定成败。模型必须在持续变化的环境中判断自己做到了哪一步,再决定继续、调整还是重来。

所以我认为,具身化的核心,就是让行动的后果进入学习过程。空间、时间、接触与动作需要形成更连贯的表示,高层推理与运动控制之间需要建立稳定的接口。直接预测动作、调用已有技能,或者交给控制器执行,都可以是这条路上的实现方式;语音与动作的协同,也会自然地融入其中。

衡量进展的标尺也随之改变:模型进入新环境,还需要多少额外经验?能完成多长的任务?出错之后,能不能自己恢复?通用能力能迁移多少,行动经验还要补多少,决定了这条路线能走多远。下面的演示能帮助直观地理解这种迁移,也能看到它还差在哪里。

Capabilities can be shared; the consequences of action have to be learned in place. Language is forgiving: say something wrong, and you can take it back in the next turn. The physical world has no undo button. A spilled cup stays spilled. Every action changes the next observation, small errors compound into later plans, and timing itself decides the outcome. A model has to keep judging how far it has gotten in a changing world, then decide whether to continue, adjust or try again.

That is why I believe the core of embodiment is letting the consequences of action into the learning process. Space, time, contact and motion need more coherent representations, and high-level reasoning needs a stable interface with motor control. Predicting actions directly, calling existing skills, or acting through a controller are all valid routes, and coordinating speech with action fits naturally into the same picture.

The yardsticks change accordingly. How much extra experience does a model need in a new environment? How long a task can it carry through? Can it recover on its own after a mistake? How much general capability transfers, and how much experience of acting must still be added, will decide how far this road goes. The demo below makes that transfer tangible, and shows what is still missing.

值得一看:Fable 5.2 把通用能力带到机械臂上

约 23 秒的演示里,同一个模型依次解魔方、拧瓶盖、装上齿轮并测试啮合,最后画了一个机器人。Anthropic 表示,能力提升也覆盖了从未被专门训练过的机器人任务;值得关注的,是同一个模型面对不同任务时的迁移能力。

也请留意画面角落的标注:2×–80× 加速,并剪去了停顿。模型已经“会做”,但还做不到现实节奏下的又快又稳。这段差距,正是具身经验要补上的部分。

另一类证据来自仿真:官方报告 Fable 5.2 在 RoboDojo 仿真基准的内部运行中平均成功率为 64.9%(同表 GPT-6 Astra 为 29.8%),超过同一测试框架下的其他前沿模型及多数机器人专用系统。这是仿真评估,不能直接当作真实机器人的成功率。

发布说明与演示 · RoboDojo 基准 · 延伸阅读:Gemini Robotics 将动作纳入模型输出

Worth watching: Fable 5.2 brings general capability to robot arms

In about 23 seconds, a single model solves a puzzle cube, unscrews a bottle cap, fits a gear and tests the mesh, then draws a robot. Anthropic says the gains extend even to robotics tasks the model was never explicitly trained on. The thing to notice is how one model carries over to very different tasks.

Also watch the labels in the corner: 2× to 80× speed-up, pauses cut. The model can already do these things, just not yet at real-world speed and steadiness. That gap is exactly what embodiment-specific experience has to close.

Separately, in internal runs on the simulated RoboDojo benchmark, Anthropic reports a 64.9% average success rate for Fable 5.2 (GPT-6 Astra scores 29.8% in the same table), ahead of the other frontier models on the same harness and most robot-specific systems. That is a simulated evaluation and should not be read as a real-robot success rate.

Release notes & demo · RoboDojo benchmark · Further reading: Gemini Robotics adds action to the model’s outputs

闭环二 · 场景Loop 2 · The scene

让真实场景成为持续学习的来源Making real settings a source of continual learning

从一个能反复验证的任务开始Start with a task you can check again and again

要回答迁移究竟发生了多少,需要一个能反复验证的现场。对大多数团队而言,从一个能持续接触的本地场景出发最有价值:任务可以被清楚定义,失败可以被完整记录,相同的条件可以用来比较不同的版本。

这里积累的不只是数据,还有对任务本身的理解:哪些偏差可以接受,什么情况下需要重新规划,怎样才算一次真正的完成。这些都要在实践中逐步明确。头部模型可以帮忙整理数据、搭建评价,但评价本身仍要与现场结果对照。只有反馈可信,训练才知道该往哪里改。

大模型从业者对此应该不陌生。编程之所以成为大模型最早成熟的应用之一,一个重要原因是代码有编译器和测试:反馈来得快,而且可信。对机器人来说,物理世界是终极的测试集,只是它运行得慢、成本高,还没法一键重置。因此,具身团队最值得做的基础建设之一,就是给机器人写“单元测试”:让“任务完成了吗”成为一个可以自动、低成本、可信地回答的问题。RoboReward 用视觉语言模型为机器人任务结果打分,就是在朝这个方向走:在真实机器人的强化学习中,它的 8B 奖励模型带来了比 Gemini Robotics-ER 1.5 更好的学习效果,并缩小了与人工奖励之间的一部分差距。

To find out how much really transfers, you need a place where you can test it again and again. For most teams, the most valuable starting point is a local setting they can keep returning to: the task can be clearly defined, failures can be fully recorded, and the same conditions can be used to compare versions.

What accumulates there is not just data but an understanding of the task itself: which deviations are acceptable, when to replan, what counts as truly done. These only become clear through practice. Frontier models can help organize data and build evaluations, but those evaluations still have to be checked against what actually happens on site. Training only knows where to go when the feedback is trustworthy.

Anyone who has worked on large models will recognize this. One big reason coding became one of the first applications of large models to mature is that code comes with compilers and tests, feedback that is fast and trustworthy. For robots, the physical world is the ultimate test suite. It just runs slowly, costs money, and cannot be reset with a keystroke. So one of the most valuable things an embodied team can build is unit tests for robots: ways to answer “did the task succeed?” automatically, cheaply and reliably. RoboReward, which uses vision-language models to score robot task outcomes, is a step in that direction. In real-robot reinforcement learning, its 8B reward model improved learning over Gemini Robotics-ER 1.5 and closed part of the gap to rewards provided by humans.

用生成的环境,扩展可以学习的经验Using generated worlds to widen experience

任务清晰之后,另一个限制会浮现:真实试错的成本高,能采集到的环境变化有限。3D 生成与重建由此有了明确的用途:围绕现场暴露的问题构建环境,改变布局、物体位置与交互条件,让机器人在回到现场之前先经历更多变化。几张图生成的空间和平面图生成的漫游,展示了构建这类环境的直观入口。

Once the task is clear, another limit appears: real-world trial and error is expensive, and the range of conditions you can collect is narrow. That gives 3D generation and reconstruction a clear job. We can build environments around the problems the site exposes, varying layouts, object positions and interaction conditions, so a robot meets far more variation before it goes back into the real world. Spaces generated from a few images and walkthroughs generated from floor plans show two intuitive entry points.

4 张概念图 → 可漫步的 3D 世界

值得关注的是,少量图像如何变成一个空间。据原帖,这套 Astra + Blender 5.3 工作流先生成 4 张概念图,把它们排布进三维空间,通过挤压、融合搭出场景,再导出到 Blender。它展示了环境构建的便捷入口;要用于物理仿真,还需要进一步加工。

imageTo3D · 约 25 秒 · 来源:X @sonic0828

4 concept images → a walkable 3D world

Watch how a handful of images becomes a space. According to the original post, this Astra + Blender 5.3 workflow generates four concept images, arranges them in 3D, extrudes and fuses them into a scene, and exports the result to Blender. It is a quick on-ramp to building environments; turning it into a physics simulation still takes more work.

imageTo3D · ~25 s · Source: X @sonic0828

2D 平面图 → 室内 3D 漫游

从房间布局生成可以走进去的三维空间,适合思考如何复用楼宇和室内场景里已有的平面资料。建议结合 World Labs 的 real-to-sim-to-real 工作一起看:空间做出来之后,怎样继续构建能用于机器人训练与评估的环境?评价环节还可以参考 RoboReward,理解如何把任务结果变成训练反馈。

to3D · 约 10 秒 · 来源:X @wanerfu

2D floor plan → 3D interior walkthrough

A room layout becomes a space you can walk into, a prompt to think about reusing the floor plans that already exist for buildings and interiors. Pair it with World Labs’ real-to-sim-to-real work: once the space exists, how do you turn it into an environment for training and evaluating robots? For the evaluation side, see RoboReward on turning task outcomes into training feedback.

to3D · ~10 s · Source: X @wanerfu

不过,生成的空间要真正服务于学习,还需要物理属性、可交互的物体、任务接口和评价标准。World Labs 的说法很有代表性:“要扩展机器人的智能,就必须扩展机器人学习所处的世界。”它报告的一项初步结果尤其关键:在其测试中,仿真里表现更好的策略,在真机上也表现更好。仿真能否给策略正确地“排序”,决定了它能否成为一个可信的评价者。

我的判断是,环境构建应当由真实的失败来牵引:现场暴露了什么问题,就补充相应的训练条件,再回到实测中检验改进是否有效。这样,仿真与真实数据才能围绕同一个目标积累,而不是各说各话。

For a generated space to serve learning, though, it needs physical properties, interactive objects, task interfaces and evaluation criteria. World Labs puts the goal well: “To scale robot intelligence, we must scale the worlds in which robots learn.” One preliminary result it reports matters most: in its tests, policies that did better in simulation also did better on hardware. Whether a simulator ranks policies correctly decides whether it can serve as a trustworthy evaluator.

My view is that environment building should be pulled by real failures. Whatever problem the site exposes, add the matching training conditions, then return to real tests to check whether the fix worked. That way simulation and real data accumulate toward the same goal instead of drifting apart.

把能力留在现场:本地部署与算力适配Keeping capability on site: local deployment and compute

这个闭环要持续转动,还依赖稳定的部署。响应太慢、网络波动、算力不足,都会改变模型实际能完成的任务。因此,部署应当与训练一起设计:哪些决策必须在机载设备上完成,哪些计算可以交给现场服务器,断网时系统还能保留哪些能力。

苹果的 HALO 提供了一个很好的参照。一台背在宇树 G1 身上的 Mac mini,约每两秒运行一次本地大模型来编排任务,每秒约三次重新生成全身动作;机器人自带的 Jetson Orin NX 则以每秒 50 次的频率跟踪执行。慢思考与快动作各在其位。部署设计的核心问题,不是模型有多快,而是哪一层闭环跑在哪里、以什么频率运行。NVIDIA 的 GR00T 2 则更进一步,把动作模型本身搬上了机器人的板载芯片。

在国内场景里,国产算力卡的适配与推理加速,值得成为团队自己的工程积累。适配需要结合目标模型、算子支持、推理框架、显存与功耗条件,在真实的任务负载下比较;部署规模扩大之后,维护方式、资源调度与长期成本也会影响方案的选择。

加速的目标应当落在任务上。量化、蒸馏、缓存复用和算子优化都可以纳入评估,但最终要检查的,是从感知到动作的端到端延迟、响应的稳定性,以及模型压缩之后任务成功率有没有下降。对机器人来说,及时做出一次可靠的决策,比孤立地追求某个推理速度指标更有意义。

For this loop to keep turning, deployment has to be stable. Slow responses, flaky networks or too little compute all change what a model can actually do. So deployment should be designed alongside training: which decisions must be made onboard, which computation can go to an on-site server, and what the system can still do when the network drops.

Apple’s HALO is a useful reference point. A Mac mini strapped to the back of a Unitree G1 runs a local large model to orchestrate the task roughly every two seconds and regenerates whole-body motion about three times a second, while the robot’s built-in Jetson Orin NX tracks that motion 50 times a second. Slow thinking and fast movement each have their place. The central question in deployment is not how fast your model is, but which loop runs where, and at what rate. NVIDIA’s GR00T 2 goes a step further and moves the action model itself onto the robot’s onboard chip.

For teams in China, adapting models to domestic AI accelerators and optimizing inference on them deserves to become core in-house engineering. Porting has to account for the target model, operator support, the inference framework, and memory and power budgets, and options should be compared under real task loads. As deployments grow, maintenance, resource scheduling and long-term cost shape the choice as well.

Acceleration should be judged by the task. Quantization, distillation, cache reuse and kernel optimization all belong in the evaluation, but what ultimately matters is end-to-end latency from perception to action, the stability of responses, and whether task success drops after compression. For a robot, making one reliable decision in time matters more than any isolated inference-speed number.

HALO:宇树 G1 + 背上的 Mac mini + 本地千问

苹果研究团队让 G1 背着一台 Mac mini,钻桌底、推箱子、侧身过门、找椅子坐下。根据论文,Mac mini M4 Pro 默认以 4-bit MLX 运行 Qwen3.6-35B-A3B 做任务编排(约每 2 秒一次),规划无需联网;全身运动生成也在 Mac mini 上完成(约每秒 3 次);底层动作跟踪则由 G1 自带的 Jetson Orin NX 以 50 Hz 执行。

值得看的,是不同层级的模型与控制如何按频率分工:慢思考,快执行。

约 46 秒 · 项目演示 · 论文 · 来源:X @QwenAIFans

HALO: Unitree G1 + a Mac mini backpack + local Qwen

Apple researchers have a G1 carry a Mac mini on its back as it crawls under a table, pushes a box, sidesteps through a doorway and finds a chair to sit on. According to the paper, a Mac mini M4 Pro runs Qwen3.6-35B-A3B (4-bit, MLX) by default to orchestrate the task about every two seconds, with no network needed for planning. Whole-body motion is generated on the same Mac mini about three times a second, and the G1’s built-in Jetson Orin NX tracks it at 50 Hz.

Notice how models and controllers at different levels split the work by rate: think slowly, act fast.

~46 s · Project demo · Paper · Source: X @QwenAIFans

GR00T 2:世界—动作模型与板载执行

可以留意各种桌面操作中双臂与灵巧手的协调。官方介绍,GR00T 2 是新一代“世界—动作模型”(WAM):同一个扩散 Transformer 同时预测未来的画面与机器人动作。其 Turbo 版本使用 2B 参数的联合视频—动作 Transformer,可在 G1 的 Jetson Thor 上运行(单次优化预测约 124 ms),并通过潜在的全身指令驱动 SONIC 全身控制器。

按官方说法,除推理模型之外,其余部分都在板载运行,因此板载执行不等于整个系统完全离线。官方标注代码将于 12 月发布。

约 91 秒 · 项目与技术报告 · Jetson Thor 平台介绍 · 来源:X @VedaAI00

GR00T 2: a world-action model, executed onboard

Watch how arms and dexterous hands coordinate across a range of tabletop tasks. According to NVIDIA, GR00T 2 is a new generation of world-action models (WAMs): a single diffusion transformer predicts future video and robot actions together. Its Turbo variant uses a smaller 2B-parameter joint video-action transformer that runs on the G1’s Jetson Thor (about 124 ms per optimized prediction) and steers the SONIC whole-body controller through latent whole-body commands.

NVIDIA says everything except the reasoning model runs onboard, so onboard execution does not mean the whole system is offline. Code is listed as coming in December.

~91 s · Project & technical report · Jetson Thor platform · Source: X @VedaAI00

当数据、环境、评价与部署彼此衔接,本地场景就能持续产出有价值的经验。团队会逐渐分清:哪些问题需要更多训练,哪些来自系统延迟,哪些已经触及控制方式乃至机体本身的极限。于是,改进的范围开始从模型延伸到整个机器人。

When data, environments, evaluation and deployment connect, a local site becomes a steady source of valuable experience. Over time, a team learns to tell which problems need more training, which come from system latency, and which have hit the limits of the control approach, or of the body itself. And so the scope of improvement begins to extend from the model to the whole robot.

闭环三 · 本体Loop 3 · The body

从改进模型,走向改进整个机器人From improving the model to improving the whole robot

把执行反馈带进控制研发Bringing execution feedback into control development

如果失败来自控制,AI 就可以参与算法编写、奖励设计、调参与测试。大模型已有的编程与推理能力,可以沿着执行记录进入研发流程:提出修改,在一致的条件下比较效果,再把有效的改进带回真实系统。早在 2023 年,Eureka 就展示了让大模型迭代编写奖励函数、帮助机器人学习灵巧技能的可能。这里最重要的积累,是让每一轮改动都能得到可追溯的反馈。

这意味着,机器人开发会越来越依赖一套能够反复运行的实验流程:问题被记录下来,候选改动在一致的条件下接受评价,仿真结果能与实测相互对照,AI 才能在连续的迭代中真正发挥作用。团队在数据组织、自动评估与开发工具上的积累,也就自然延伸到了运控研发。

If a failure comes from control, AI can take part in writing algorithms, designing rewards, tuning and testing. The coding and reasoning abilities of large models can follow the execution logs into the development process: propose a change, compare it under consistent conditions, and carry the improvements that work back to the real system. As early as 2023, Eureka showed that a large model could iteratively write reward functions that help robots learn dexterous skills. The most important thing to accumulate here is traceable feedback for every change.

This means robot development will rely more and more on an experimental pipeline that can run again and again: problems get recorded, candidate changes are evaluated under consistent conditions, and simulation results can be checked against real tests. Only then can AI contribute across continuous iterations. A team’s investment in data organization, automated evaluation and developer tooling extends naturally into motion-control research.

当控制遇到边界,让身体也参与优化When control hits its limits, let the body join in

但控制算法的空间受身体约束。关节配置、连杆尺寸、执行器与传感器的布局,决定了机器人能看到什么、能做出怎样的动作。如果机械设计与控制研发长期分离,很多问题只能在机体定型之后,由算法一遍遍地补偿。

更进一步的可能,是让任务表现同时塑造控制策略与身体设计。当模型能够在不同机体之间学习动作规律,本体就有机会成为一个可以一起调整的变量:对某类任务更合适的结构,可能降低控制的难度;更强的控制能力,也可能让新的结构方案变得可行。下面两段视频提供了观察这种关系的视角。

But the space of control algorithms is bounded by the body. Joint configuration, link lengths, and the placement of actuators and sensors determine what a robot can see and which motions it can make. When mechanical design and control development stay separate, many problems can only be patched by algorithms, over and over, after the body is frozen.

A further possibility is to let task performance shape both the control policy and the body design. Once models can learn motion patterns across different bodies, the body itself becomes a variable that can be tuned along with control. A structure better suited to a class of tasks can make control easier; stronger control can make new structures feasible. The two videos below offer a way to look at this relationship.

Transformer Transformer:让身体也成为设计变量

这段约 28 秒的演示,适合对照目标运动观察机器人结构的变化。Huy Ha、Karen Liu 与 Shuran Song 在 CoRL 2026 论文中把本体生成与跨本体控制放进同一个模型:给定目标运动与奖励,它生成一台完整的机器人,包括每根连杆、每个关节与电机以及惯性参数,再为这副身体生成控制动作。

团队还实际制造并测试了优化版 ALOHA:跟踪误差较原设计降低 73%,最大关节速度降低 30%。这体现了协同设计的潜力,但还不意味着可以为任意任务自动找到最优的机体。

项目 · 论文 · 代码 · 视频来源:X @VedaAI00

Transformer Transformer: the body becomes a design variable

Watch this ~28-second demo with the target motion in mind and see how the robot’s structure changes. In their CoRL 2026 paper, Huy Ha, Karen Liu and Shuran Song put body generation and cross-embodiment control into a single model: given a target motion and a reward, it generates a complete robot, down to every link, joint, motor and inertial property, and then generates the actions that control that body.

The team also built and tested an optimized ALOHA, reporting 73% lower tracking error and 30% lower maximum joint speed than the original design. That shows the promise of co-design; it does not yet mean we can automatically find the best body for any task.

Project · Paper · Code · Video source: X @VedaAI00

从结构与运动看本体设计

约 6 秒的短片,适合观察连杆、关节与运动方式之间的关系。与旁边的协同设计演示放在一起,可以进一步思考:哪些动作上的困难来自控制,哪些来自身体结构?具体方法以原帖为准。

视频来源:X @VedaAI00 · 延伸阅读:Eureka:用大模型迭代奖励函数,辅助控制技能学习 · 自动机器人设计的早期研究

Reading body design through structure and motion

A ~6-second clip for looking at how links, joints and movement relate. Next to the co-design demo, it invites a further question: which difficulties in a motion come from control, and which from the body’s structure? See the original post for the method.

Video source: X @VedaAI00 · Further reading: Eureka: LLMs iteratively writing reward functions for skill learning · Early work on automatic robot design

从一次协同设计,走向持续的共同演进From one-off co-design to continual co-evolution

沿着这个方向,我期待的是软硬件的共同演进(co-evolution):现场执行暴露问题,模型提出控制或本体的修改,仿真帮助筛选方案,制造与实测再提供下一轮反馈。评价也要纳入能耗、耐用性、成本与维护,让每一次改进都更贴近长期使用的要求。

这条链路上的环节,正在被逐一打通。Anthropic 报告,Fable 5.2 与电路板设计制造公司 Diode 合作,从一段需求描述出发,完成了三款电子设备的规格制定、元件选型、四层电路板布线、外壳建模与软件编写,三款设备都无需任何电气修改就能工作。模型能参与的环节,正从写代码延伸到造实物。

这也是物理世界递归自我改进(RSI)的一条可能路径。要让它持续运转,设计、制造、测量与评估都必须接入同一个研发过程。它的长期吸引力在于:AI 获得的经验,最终可以反过来改善承载它自身能力的系统。

回头看,自然界中的大脑与身体,是在数亿年里一起演化出来的;机器人领域却长期把二者分开设计:先定下身体,再让算法去适应。进化机器人学很早就尝试过让形态与控制一起演化。如今不同的是,通用模型带来了理解任务、编写代码和完成设计的能力,生成环境与快速制造则缩短了每一轮迭代的周期。大脑与身体,有机会重新一起演进,而且不再以地质年代为单位,而是以一轮轮工程迭代为单位。

Following this direction, what I look forward to is the co-evolution of software and hardware: field execution exposes problems, models propose changes to the control or the body, simulation screens the candidates, and fabrication and testing supply the next round of feedback. Evaluation should also weigh energy use, durability, cost and maintenance, so that each improvement moves closer to what long-term use demands.

The links in this chain are being connected one by one. Anthropic reports that, working with Diode, a circuit-board design and manufacturing company, Fable 5.2 took a one-paragraph request and delivered three electronic devices, covering specs, parts selection, four-layer board layouts, CAD enclosures and software, and all three worked without any electrical modifications. What models can take part in is extending from writing code to making things.

This is one possible path to recursive self-improvement (RSI) in the physical world. To keep it running, design, fabrication, measurement and evaluation all have to plug into the same development process. Its long-term appeal is that the experience AI gains can ultimately improve the very system that carries its capabilities.

In nature, brains and bodies evolved together over hundreds of millions of years. Robotics has mostly designed them apart: fix the body first, then make the algorithms cope. Evolutionary robotics explored co-evolving form and control long ago. What is different now is that general models bring the ability to understand tasks, write code and produce designs, while generated environments and rapid fabrication shorten every round. Brain and body have a chance to evolve together again, measured not in geological time but in engineering iterations.

落点 · 从哪里开始Next · Where to start

把能力增长接到自己的实践中Plugging the growth of capability into your own practice

回到开头的问题:为什么是现在?因为通用模型正在提供越来越高的能力起点,而具体场景中的数据、评价与部署体系,仍然需要亲手建立。越早把模型接入真实任务,就越早开始积累只属于自己的经验,并在下一代模型到来时,验证、吸收和放大它的能力。

So, why now? Because general models are providing an ever-higher starting point, while the data, evaluation and deployment systems of a specific setting still have to be built by hand. The sooner you connect models to real tasks, the sooner you start accumulating experience that is yours alone, and the better placed you are to validate, absorb and amplify the next generation of models when it arrives.

如果明天就开始If you start tomorrow

  1. 选一个每周都能接触到的真实场景,写清楚任务,以及什么才算“完成”。
  2. 记录每一次失败,连同复现它所需的条件。
  3. 先为任务写“单元测试”:一个廉价、可信的成功判定,往往比一个更大的模型更早产生价值。
  4. 让仿真从真实的失败中生长出来,并用实测检验每一次改动。
  5. 想清楚哪一层闭环跑在哪里,端到端地测量从感知到动作的延迟。
  6. 新模型发布的那天,把整套评测重跑一遍。你的闭环,就是吸收下一代能力的接口。
  1. Pick a real setting you can access every week. Write down the task, and what counts as done.
  2. Log every failure, along with the conditions needed to reproduce it.
  3. Write unit tests for the task before anything else: a cheap, trustworthy success check often pays off sooner than a bigger model.
  4. Grow your simulation out of real failures, and verify every change on hardware.
  5. Decide which loop runs where, and measure latency end to end, from perception to action.
  6. The day a new model ships, rerun your whole evaluation suite. Your loop is the interface through which you absorb the next generation.

对大模型与 AI 从业者来说,这个窗口值得认真把握。从一个可以持续接触的场景开始,把熟悉的模型能力接到行动上,再用真实的反馈推动模型、控制与硬件一起改进。通用模型打开了进入具身的大门;持续的实践,则让这份通用能力逐渐长成你自己的机器人能力。

For people working on large models and AI, this window deserves to be taken seriously. Start with a setting you can keep coming back to, connect the model capabilities you know well to action, and let real feedback drive the model, the control and the hardware to improve together. General models have opened the door to embodiment; sustained practice is what turns that general capability into robot capability of your own.

大模型让机器读懂了世界;具身智能,要让它在世界中学会行动,并在一次次行动里,与自己的身体一起成长。Large models taught machines to read the world. Embodied AI will teach them to act in it, and, one action at a time, to grow together with their bodies.

参考资料References

  1. Fortune. Exclusive: OpenAI disbands its stand-alone robotics team as its general models start running robots. 2026.
  2. Google DeepMind. Gemini Robotics brings AI into the physical world. 2025.
  3. Anthropic. Introducing Claude Fable 5.2 and Claude Mythos 5.2. 2026. See also the RoboDojo benchmark.
  4. Lee, Wagenmaker, Pertsch, Liang, Levine & Finn. RoboReward: General-Purpose Vision-Language Reward Models for Robotics. arXiv:2601.00675, 2026.
  5. World Labs. Building Worlds That Train Robots. 2026.
  6. Hu et al. (Apple). Perceptive Humanoid Motion from Simulation (HALO). arXiv:2611.06908, 2026. Project page.
  7. NVIDIA. GR00T 2: Scaling Robot Foundation Models for Generalization, Dexterity, and Adaptation. 2026.
  8. NVIDIA. Introducing NVIDIA Jetson Thor, the Ultimate Platform for Physical AI. 2025.
  9. Ma et al. Eureka: Human-Level Reward Design via Coding Large Language Models. 2023.
  10. Ha, Liu & Song. Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design. CoRL 2026. Project · Code.
  11. Matthews et al. Efficient automatic design of robots. PNAS, 2023.