新智元报道
一块普普通通的面包，变换了摆放姿态，常常就能难倒「已经学会了」的机器人！
训练时，面包通常是竖放的；
如果把面包横过来，机械臂明明还能识别、也会抓取，却总是偏了那么一点距离，被难倒在了这个幼儿园小孩都能轻松解决的问题上。。。
在这个「横放面包」的测试设置里，原始策略模型的成功率只有 5%。
它并不是不会。抓取、移动、放置，整套技能都在，只是这次的初始位姿越出了训练分布（OOD，Out-of-Distribution），导致它「做偏了」。
重新采集示教数据、增加训练吗？不仅需要额外时间投入，对算力也是不小的负担。
更麻烦的是，端到端模型的「黑盒」特性让人很难真正解释，到底是哪一步出了问题？
好消息是，面对「已有任务能力、但某一步做偏」的情况，极智嘉（Geek+）具身智能最新给出的答案是——不用回炉，用大白话即修正机器人局部动作偏差！
近日，极智嘉在其 Gravity 具身智能框架下推出 Gravity Control 因果干预模型。
它深入黑箱模型内部，找到真正「握着方向盘」的那些特征，在推理阶段直接调节。
面对分布外任务，操作人员只需给出一句自然语言提示（如「抓面包的时候更深一些。」），机器人就能把跑偏的动作改过来。
结果是：在真实的分布外复杂操作任务测试中，整体成功率从 17.5% 大幅提升至 87.5%；其中横放面包这一单项 OOD 设置，直接由 5% 飙升至 95%！
这组惊艳的数字背后，是一个值得全行业关注的变化——
机器人纠错这件事，第一次有了一条可解释、可验证、可约束的运行时路径！
过去两年，具身智能模型的进步有目共睹，但是真实物理世界从不按训练集出牌。
大量失败并不来自「完全不会」，而来自「差一点」。
这类局部偏差最折磨人。
对部署方而言，它意味着：现场不能停，问题又说不清，只能把数据上传回去，等下一版模型。
于是问题变成：对于模型已有能力范围之内的执行偏差，有没有一种更直接、可验证的纠偏方式？
要在推理运行时纠偏，先得回答一个底层问题：模型内部成千上万的高维隐藏表征里，哪些才真正掌握动作的控制权？
学界流传着一个迷思：如果某个表征能被解码出「抓取」「向下」这类语义概念，它就一定在指导机器人的行为。
但极智嘉研究团队的结论给这个直觉泼了一盆冷水——相关性不等于因果性。
研究发现，许多对物体或指令高度敏感的语言层表征，扮演的正是「解说员」——它对机械臂怎么动，几乎没有实际贡献。
真正驱动末端执行器姿态、夹爪开合与位移速度的「球员」，高度集中在动作生成链路深处的动作专家层。
为厘清作用路径，极智嘉团队形式化定义了一条概念属性链，一个特征要依次通过五道关卡：
Faithful（保真性） → Motor Aligned（运动对齐） → Verified（泛化验证） → Causal（因果可控） → Safe（闭环安全）
实验结果揭示了一个颠覆性的现象：语言层特征往往在「运动对齐」或「泛化验证」环节就从传导链上断开了！
名字听上去与运动有关，真去干预，物理动作却纹丝不动。只有动作专家层的特征，能完整走完全部五关。
这一发现的分量在于，它把「可解释」和「可控制」彻底拆开了。
不少可解释性工作停留在「给特征命名」，而 Gravity Control 坚持一条更严格的标准：一个特征有没有控制权，不应该看它叫什么，而更应该看干预之后动作有没有按预期响应。
基于上述发现，极智嘉构建了全新的 Gravity Control 模型。
Gravity Control 模型框架总览
它的工作方式可以概括为精妙的三步：
第一步，拆（稀疏特征发现）。在冻结原始模型骨干的前提下，于残差流中引入 Top-K SAE（稀疏自编码器）等模块，把高维隐藏激活做稀疏分解，提取出一个个具有人类可解释语义的 Latent Feature。
第二步，推（定向因果干预）。对已完成语义标注的特征做定向因果干预（放大或抑制），再观察位置、旋转、夹爪等动作通道的真实物理响应。
第三步，筛（特异性评估）。引入干预特异性评估，只保留方向一致、剂量响应单调、影响集中在目标动作上的特征。通俗地说：调深度就只动深度，不能顺手把手腕也拧了。通过这三步的，才是真正可用的控制手柄。
π0.5 与 OpenVLA-OFT 在不同层插入 SAE 后得到的可用特征数量比例区别
相较于过去业内只能通过特征激活对模型做被动「读出」，Gravity Control 真正打通了 Interpretability → Intervention → Control 的完整闭环。
这，就是用白盒化、可解释的思路来做因果干预。
找到了控制手柄之后，还有一个更具工程意义的问题：人怎么用最方便？
当然是说人话了！
传统端到端模型，只接受「把面包放进盘子」这样的宏观任务级指令。
但宏观指令并不天然提供对抓取深度、手腕角度等局部动作的稳定控制。
而 Gravity Control 会基于已建立的概念控制表，搭建一条语言引导的动作调节链路。
当机器人出现偏差，操作人员发出一句自然语言提示，外部 VLM 规划器无需重新规划整条轨迹，而是检索概念表中的特征词表，直接对原始模型内部激活做在线调节。
与此同时，系统引入了闭环门控，为干预程度设置上限，确保调节只作用于特定几何自由度，不会因为「用力过猛」引发灾难性的动作发散。
换言之，上一步骤的调整，不会影响到原本就能准确完成的下一步动作。
让机器人根据自然语言指令执行任务，已经是多种 VLA 模型具备的能力。
Gravity Control 最关键的进步是：当机器人已经具备完成任务的基础能力，却在执行中出现偏差时，如何把「抓得更深一点」这样的纠偏提示，转化为模型内部可验证、可约束的动作调节，在不更新原始策略权重的前提下修正偏差。
把视线拉到当前百家争鸣的全球具身智能赛道，会发现一个有趣的现象——
绝大多数头部玩家的主战场，是「让机器人会得更多」；
而「已经会的事做偏了怎么办」，至今仍是一块少有人系统攻克的硬骨头。
第一类：能力派——不断外推「会做什么」的边界
这一派代表了当前具身大模型的主流叙事。
Figure AI 在 2026 年初推出 Helix 02，随后展示了机器人在完全自主控制下连续跑满 8 小时班次的能力；其 Index 预训练将零样本成功率从约 9% 提升至 56%。
谷歌 DeepMind 发布的 Gemini Robotics 2 与 RT 系列，主打全身控制与多机协作。
英伟达开源了 30 亿参数的 Isaac GR00T，提供面向不同机器人的微调工作流；
Skild AI 则以「一个大脑控制任意机器人」为目标疯狂吸金。
这些工作共同把能力边界向显著拓展。
但当具体部署中出现局部偏差，主流的修复手段仍然是「黑盒模式」：回到数据，重新微调——以 GR00T 为例，其标准落地路径本身就是「基础模型 + 场景微调」。
第二类：数据派——用海量真实部署换可靠性
这一派，以特斯拉 Optimus 最有代表性。
马斯克曾坦言，工厂里还没有 Optimus 在做真正有用的工作，直到近期产量提升，精细操作依然是难题。
他们高度依赖庞大的人类遥操作团队，一旦遇到新环境表现不佳，做法就是「堆数据」重新训练。
同样的，Agility Robotics 的 Digit 通过累计超 6.5 万小时的运行将数据反哺模型；
Apptronik 建立巨大测试场专门采集数据。
这条路有效，但极其昂贵——可靠性的提升以「月」为单位，以「万小时」计。
第三类：纠错派——探索「失败恢复」的新范式
这是与 Gravity Control 愿景最接近的领域。
Physical Intelligence 的 π0.6 提出了 RECAP 方法，把示教、专家纠正与自主经验结合进强化学习，是「从错误中学习」的标杆。但其逻辑是「纠正 → 回收数据 → 再训练」，纠错依然发生在训练回路里。
1X Technologies 的 NEO 采用「Expert Mode」：遇到不会的任务由远程人类专家 VR 接管，本质是「人替机器人干」。
学术界（如斯坦福、伯克利等校）开始尝试通过 VLA 内部表征直接引导机器人行为，证明了「不改权重也能调行为」在技术上可行。
这样一看，极智嘉 Gravity Control 的领先性与独特生态位就非常清晰了。
它实现了四个「同时做到」——
1.零权重更新、零新增数据：相比训练回路纠错，将纠偏周期从「天/周」压缩到「一句话」。
2.完整的因果验证体系：通过概念属性链 + 特异性评估，把学术界「能调」的探索，推进到了「可控且安全」的工程落地标准。
3.自然语言接口与闭环门控：不需要操作员戴上 VR 手把手示范，同时有剂量上限兜住安全边界。
4.真机 OOD 任务的量级提升：单项成功率从 5% → 95% 的跃升，在公开披露的同类运行时干预成果中极其罕见。
更值得注意的是，这项技术的问题定义来自真实的产业作业，而非实验室。
目前，极智嘉的业务已覆盖了 40 多个国家、超过 1700 个项目，其面对的是仓储拣选、多 SKU、光照变化等高频场景中的极致可靠性要求。
在具身智能从 Demo 走向生产力的阶段，Gravity 框架补上的，正是解决「落地部署后迭代成本高昂」这块拼图。
这套技术不仅成绩亮眼，更实在的一点是，它完全不绑定特定的底层模型。
目前，它已在 π0.5、OpenVLA-OFT 等代表性 VLA 架构、LIBERO 仿真基准以及物理机器人平台上完成评估。
概念命名与动作因果控制相脱节 Naming and Control Come Apart 现象
看一眼实测数据，就知道这有多大的提升——在真实的物理机器人 OOD 任务中，在完全不更新原始策略权重、不新增示教数据的前提下，Gravity Control 将横放面包等分布外失败任务的整体成功率从 17.5% 跃升至 87.5%。
特别是前面提到的「横放面包」测试，成功率从 5% 提升到了 95%。
利用自然语言在线纠偏真实机器人的完整工作流
如果沿用补采数据、重新微调的办法来争取同样的提升，还得投入多少训练算力和迭代时间？而这次纠偏，是在不更新原始策略权重的前提下实现的。
当然，也得客观看待，这套方案不是什么「无中生有」的魔法。它的定位很明确：是「纠偏」而不是「启蒙」。
前提是机器人本身得具备这项基础技能，只是因为现场的一点变化（比如东西放歪了）导致动作跑偏了，它才能把偏掉的动作给「拉回来」。
想靠它直接教会机器人一门全新的手艺是不现实的。
但如果你在一线做过真实的商业部署，就会深有体会：恰恰就是这种「明明会干，但换个姿势就死活干不好」的微小偏差，才是每天在现场最耗费人力的无底洞。
能把这个问题低成本地解决掉，本身就已经解决了一大半的落地难题。
回到真实的柔性制造或仓储现场，Gravity Control 的价值会被放得无限大。
一个工作站每天面对成百上千种 SKU，摆放随机、包装各异。哪怕只有 1% 的动作跑偏，落到日均数千次抓取上，也是数十次人工介入。对客户而言，真正的成本在部署之后的每一天。
Gravity Control 指向的，是两层更长期的系统价值：
第一层，自动纠错闭环。
未来，Gravity Control 可以接入上层的 Agent 或 AI 原生系统套件。由系统直接充当「监工」，实时检测（Detect）任务状态 → 诊断（Diagnose）动作偏离 → 主动调用 Gravity Control 纠正（Correct）。
第二层，经验沉淀与复用。
一次成功的纠偏绝非「阅后即焚」。系统可以记录当时的任务阶段、场景状态、被调用的 Control Handle 以及干预剂量，把它沉淀为可复用的 Skill。
从纠错到 Skill，再到自进化——Gravity Control 为具身系统走向自主纠错与能力积累，提供了一条清晰的工程路径。
现在，具身智能的能力曲线仍在快速上扬，但真实世界不会只奖励「会做」，它给你一个更棘手的问题：「做错了怎么办？」
Gravity Control 给出的回答是：当模型已经具备基础能力时，就不必每次都回到昂贵的数据工厂。
只要找到真正掌握控制权的内部特征，用一句自然语言在推理阶段精准调节，就能用闭环门控守住安全边界，再把每一次成功纠偏沉淀为经验。
对研究者而言，这证明了「概念可解释 ≠ 行为可控制」的实证；而对于产业而言，这意味着一笔实实在在的账：机器人在现场部署后的运维和微调成本，终于有希望降下来了。它不再是个需要人天天盯着的娇贵设备。
分布外复杂操作任务的的成功率从 5% 拔高到 95%，可能只是一盘前菜。真正让人兴奋的是，给机器人「纠错」这件原本极其昂贵、费时又像开盲盒一样的事，终于变成了一种看得懂、管得住、还能不断积累的能力。
怎么让具身智能从酷炫的 Demo 变成真正好用的生产力工具？在这个问题上，极智嘉已经率先交出了答卷。
编辑：Aeneas 马可
不喜欢
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。
New Ji-won.
A piece of ordinary bread, with a change of posture, can often be reduced to a robot who has learned!
When trained, bread is usually erected;
If the bread is crossed, the arm of the machine can be identified and captured, it is always a little off, and it is hard to fall on the children of this kindergarten who can easily solve the problem.
success rate is only 5% of the original policy model in this "bread out" test setting.
It's not gonna happen. The capture, movement, placement, and skill sets are all in place, except that the initial position of training (OOD, Out-of-Distribution) went beyond training and led it to "wrong".
Recollection of teaching data and additional training? Not only would additional time be required, but it would also be a considerable burden on the economy.
Even more troublesome, the end-to-end model's "black box" feature makes it difficult to really explain.
The good news is that, in the case of "a mission that has the capacity but one step is biased", Geek+'s smart and recent answer is -- to correct the robot's local motion deviations with big white words without turning back.
Recently, within the framework of Gravity ' s smartness, the Gravity Control Intervention Model was launched.
It goes deep inside the black box model, finds those features that really hold the wheel, and regulates them directly at the reasoning stage.
In the face of an undistributed task, the operator simply needs to give a natural language hint (e.g., "Deeper when catching bread.") so that the robot can change the running motion.
As a result, the total success rate rose from 17.5 per cent to 87.5 per cent in the real out-of-distribution complex operation test; the single OOD set-up, which spreads bread, rose directly from 5 per cent to 95 per cent!
Behind this amazing number is a change that deserves industry attention.
For the first time, the robot has an interpretable, verifiable and binding running path!
Advances in smart models have been visible over the past two years, but the real physical world has never been trained to play cards.
Many of the failures came not from "never" but from "nearly".
These partial deviations are the worst.
For the deploying party, it means that the site cannot be stopped and the problem is not clear, but that the data can be uploaded back to the next version of the model.
The question then becomes: Is there a more direct and verifiable way to correct the deviations in the implementation of the model within its capabilities?
In order to correct the reasoning, one of the bottom questions must be answered: What is really in control of the movement?
There's a myth that if a sign can be decoded as a semantic concept such as "capturing down," it must be guiding robots.
But the conclusions of the Ultra-Intelligence Research Team are pouring cold water into this intuition — relevance is not the same as causality.
The study found that many highly sensitive linguistic layers of objects or commands play precisely the role of `interpreters' — how they move a mechanical arm with little practical contribution.
The "player" that really drives the end implementer's attitude, clamps up and move speeds, is highly concentrated in action specialists deep in the action generation chain.
In order to clarify the path to action, the SWAT formally defined a conceptual attribute chain, with one feature passing in turn through five levels:
Faithful → Motor Alined → Verified → Causal → Safe
The results of the experiment reveal a subversive phenomenon: the language layer features are often cut off from the chain of transmission by the "motion alignment" or "penetration" link.
It sounds like it's sports-related, it's interference, it's physical action. Only the signature of the action specialist layer can complete the full five.
The point of this discovery is that it completely disintegrates "explainable" and "controllable".
A lot of explanatory work has been left in the "naming of features", and Gravity Control insists on a more stringent criterion: a feature has no control, not its name, but rather whether the post-intervention action responds as expected.
Based on the above findings, a new Gravity Control model has been constructed.
Gravity Control Model Frame Overview
Its working methods can be summed up as three fine steps:
First step, disassembly. In the context of freezing the original model backbone, modules such as Top-K SAE (the thin coder) are introduced into the residual stream, where the cavity is activated by thin decomposition, extracting a human-explainable Latin Feature.
Step two, push (directive cause and effect intervention). Directed cause-effect intervention (magnification or inhibition) with completed semantic labels, followed by a real physical response to action channels such as positions, rotations, clamps, etc.
Step three, screening (specific assessment). Introduce an assessment of the specificity of the intervention, retaining only characteristics that are consistent in direction, dosage-specific, impact focused on target actions. It's popularly said: the Depth is only moving, and it can't be twisted by step by step's wrist. It's through these three steps that the control handle is really available.
The ratio of thorium 0.5 to the number of available features obtained by OpenVLA-OFT inserting SAE on different layers
Compared to the past industry, where passive "read-out" of models can only be done through characterization activity, Gravity Control actually opens the full closed loop of the Interpretability → Intervention → Control.
This is the case with a white boxing, interpretable approach to causal intervention.
Once the handle was found, there was a more engineering question: how can people use it most easily?
Of course it's human!
Traditional end-to-end models are only subject to macro-level mission-level instructions such as "Put bread in the plate".
However, macro-directions do not naturally provide stable control over local movements such as capture depth, wrist angles, etc.
And Gravity Control will build a language-led movement-regulating link based on a built concept control sheet.
When robots deviate, the operator sends a natural language hint that the external VLM planner does not need to reprogram the entire trajectory, but rather to retrieve the characterizations in the concept sheet and directly adjust the internal activation of the original model online.
At the same time, the system introduced closed-ring door control, with a ceiling on the level of intervention, to ensure that the reconciliation functioned only at a given geometric level of freedom and did not cause catastrophic movements by “too hard”.
In other words, the adjustment of the previous step would not affect next step actions that could have been completed accurately.
Making robots operate according to natural language instructions is already the capability of multiple VLA models.
Gravity Control is the most important step forward: how to translate the "tweak deeper" type of bias when robots have the basic ability to perform their tasks, and when there is a bias in their implementation, into a model-based, verifiable and binding action adjustment that corrects the bias without updating the original policy weight.
It's an interesting phenomenon to pull the sights to the current 100-challenging global smart track.
The main battleground for most head players is "Let the robot get more."
And "what if we do what we've already done?" is still a hard bone that is being systematically attacked by fewer people.
Type one: Capable pie -- constantly pushing the "what" border
It represents the current mainstream narrative of LLM.
Figure AI launched Helix 02 in early 2026, and then demonstrated the ability of robots to run for eight consecutive hours under full self-control; its Index pre-training increased zero Q111success rate from about 9% to 56%.
Gemini Robotics 2 of DeepMind Z94Z Google and the RT series, with full-body control and multi-machine collaboration.
Isaac GR00T, with 3 billion parameters, provides fine-tuning workflow for different robots;
Skild AI is targeting "a brain-controlled robot" for the wild use of gold.
Together, these efforts have significantly expanded the capacity boundary.
But when partial deviations occur in specific deployments, the mainstream means of repair remain the "black box model": back to the data, fine-tuned — for example, the standard landing path itself is the "basic model + scenario fine tune."
CATEGORY II: DATA DISTRIBUTION — RELIGIOUS TRANSPORT FOR RELIEF
This is the most representative faction of Tesla Optimus.
Mask once said that Optimus was not doing really useful work at the factory, and until the recent increase in production, precision remained a challenge.
They are highly dependent on large human teleworking teams, and once new environments are underperforming, the practice is to retrain "stack data".
Similarly, the Digit of Agility Robotics uses a cumulative run of over 65,000 hours to refeat the data;
Apptronik creates huge test sites dedicated to collecting data.
The road works, but it's extremely expensive — reliability is raised in months, in 10,000 hours.
Category III: Transcripts — exploring a new paradigm of “failure recovery”
This is the closest area to the Gravity Control vision.
Physical Intelligence 0.6 proposes an RECAP approach that combines teaching, expert correction and autonomous experience with enhanced learning, which is the "learning from mistakes". But the logic is "correcting data recovery, retraining," and the error still occurs in the training back road.
The NEO of the 1X Technologies uses the term "Expert Mode": tasks that are not likely to happen are taken over by a remote human expert, VR, which is essentially a "man-made robot."
Academics (e.g. Stanford, Berkeley, etc.) have begun to try to direct robotic behaviour through internal VLA signs, proving that "no change in weight can be made" is technically feasible.
In this way, the leading and unique ecological position of Gravity Control is very clear.
It achieves four "do it at the same time":
1. Zero weight updates, zero new data: the correction cycle is reduced from "day/week" to "said" compared to the training circuit.
A complete causal certification system: the "mutual " search of academia through the conceptual attribute chain + specificity assessment has advanced to the "controlled and safe " engineering landing standard.
3. Natural language interfaces and closed loop door controls: no operator is required to wear VR step by step demonstrations, with dose ceilings to cover safe borders.
Real machine OOD task scale upgrades: individual success rate leaps from 5% to 95%, extremely rare in publicly disclosed type of operational intervention.
More notably, the definition of the problem of this technology comes from real industrial operations, not laboratories.
At present, the ultra-wise operations have covered more than 40 countries and more than 1700 projects, facing extreme reliability requirements in high-frequency scenarios such as warehousing selection, multiple SKU and photochange.
The Gravity framework was added to the intellectual transition from Demo to productivity by addressing the puzzle of the "high cost of deployment after landing."
Not only is the technology promising, but it does not bind specific bottom models at all.
At present, it has completed its assessment on representative VLA structures, LIBERO simulation benchmarks and physical robotic platforms.
Naming and Control Come Part
Looking at the data, you will see how high it is. In the real physical robot OOD mission, Gravity Control jumped from 17.5 per cent to 87.5 per cent the whole success rate of the out-of-distribution failed tasks, such as bread, without updating the original strategy weight at all and without adding teaching data.
Especially with the "bread out" test mentioned earlier, success rate rose from 5% to 95%.
Full workflow of real robots online in natural languages
How much more time does it take to train the calculus and integrator if the same upgrade is to be achieved by using complementary and fine-tuning methods? And this is done without updating the original strategic weight.
And, of course, it has to be seen objectively, it's not some kind of magic that's out there. It's well-positioned: it's "repeated" and not "enlightened."
If the robot itself has to have this basic skill, it can't pull it back if it's lost action because of a slight change in the field (e.g., a misdirection).
It's not realistic to teach robots a whole new kind of work.
But if you do a real business deployment at the front line, you know it: it's exactly the kind of small deviation that "we can do it, but we can't do it in a different position" and it's the most labour-intensive bottom hole on the ground every day.
By addressing this problem in a low-cost manner, it has resolved half of the problem of landing.
Back to real soft manufacturing or storage, the value of Gravity Control would be unlimited.
A workstation faces hundreds of SKUs each day, with random packages. Even if only 1% of the movement runs, thousands of times by day, it's dozens of manual interventions. For clients, the real cost is every day after deployment.
Gravity Control points to two layers of longer-term system value:
First floor, automatically correct the loop.
In the future, Gravity Control can access top-level Agent or AI primary system packages. Directly as a "supervisor" by the system, real-time testing of the Diagnose mission is deviating from the initiative to call Gravity Control Correct.
Second tier, empirical deposition and reuse.
A successful correction is not to "burn it after reading." The system can record the mission phase at that time, the state of the scene, the callable Control Handle and the intervention dose, and deposit it into a reusable Skill.
From error correction to Skyll, to self-evolution — Gravity Control — provides a clear engineering path to self-censorship and capacity-building.
Now, the smart curve is still rising fast, but the real world will not be rewarded only for "do it," which gives you a more difficult question: "What about wrong?"
Gravity Control replied that when models are already basic, they do not have to go back to expensive data factories each time.
As long as the internal features of true control are found, and the natural language is fine-tuned at the reasoning stage, secure borders can be secured with closed loops, and each success can be reduced to experience.
For researchers, this is proof that "the concept explains how behavior can be controlled"; and for industry, it means a real account: there's hope that the peacekeeping and fine-tuning costs of robots after their deployment on the ground will come down. It's no longer a precious device that needs to be watched every day.
The success rate, which is distributed outside complex operations, is up from 5% to 95%, probably just a plate of front-runners. What's really exciting is that the "correcting" of robots, which was so expensive and time-consuming, turned into an ability to understand, manage and build up.
How do you make smart Demo a really good productivity tool? On this issue, Qui Chia has taken the lead in handing over the responses.
Editor: Aeneas Marco
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.