具身智能公司辛苦训练的机器人大脑，要被GPT、Claude、Gemini抢饭碗了？10月14日，知名大模型评测机构Artificial Analysis把这些在数学、编程等领域表现出色的大模型，拉进了一场真正的机器人考试。在新发布的AA-Robotics v0.1评测中，GPT、Claude、Gemini等前沿模型需要控制真实机械臂，挑战倒水、写字、搭积木等八项任务，比一比谁的动手能力更强。
Will the robot brains that embodied intelligence companies painstakingly trained be replaced by GPT, Claude, and Gemini? On October 14, well-known large model evaluation organization Artificial Analysis pulled these models that excel in math, programming, and other fields into a real robot exam. In the newly released AA-Robotics v0.1 evaluation, frontier models including GPT, Claude, and Gemini must control real robotic arms, tackling eight tasks including pouring water, writing, and building block towers, competing on whose "hands-on ability" is stronger.
结果，大模型们刚一上手，就闹出了不少笑话。一台机械臂花了十几分钟，终于把几块泡沫积木搭成了一座合格的高塔。背后指挥它的GPT-6.1 Sol明明已经完成任务，却仍然不肯停手，继续调整积木，结果反而让这项任务以失败告终。这只是此次机器人考试中颇具戏剧性的一幕。首批结果同样充满反差，GPT-6 Astra排名第一，通用大模型已经能够借助编程Agent操控真实机械臂，但它们完成一项任务平均需要约半小时，比人类基线慢了将近两个数量级。
The result: the large models had barely gotten started before producing plenty of comedy. One robotic arm spent over ten minutes finally building several foam blocks into a qualified tower. The GPT-6.1 Sol behind it had clearly completed the task but refused to stop, continuing to adjust blocks, ultimately causing the task to end in failure. This is just one dramatic scene from this robot exam. The first batch of results is equally full of contrast—GPT-6 Astra ranks first, general large models can already control real robotic arms via coding Agents, but they take about half an hour on average to complete one task, nearly two orders of magnitude slower than the human baseline.
这份成绩单，也让机器人行业最近的一场争论有了新的讨论依据。随着通用大模型越来越强，具身智能公司专门训练的机器人大脑，究竟还有多少不可替代的价值。上个月，GPT-6 Astra控制机械臂画金门大桥、抓取积木、模仿人类操作的视频接连刷屏。一台售价约150美元的机械臂，在Astra的控制下，竟然能根据摄像头画面自主规划动作，一笔笔画出金门大桥。更有说服力的实验来自RoboCurve，在它的实机测试中，Astra控制双机械臂，把红色积木放进碗里，20次成功了19次。这样的成绩，让不少人开始追问，既然通用大模型已经懂视觉、空间、推理，还能编写控制代码，机器人公司为什么还要投入大量时间和金钱，从头训练自己的VLA模型。有人甚至提出，机器人领域的OpenAI，可能就是OpenAI自己。
This report card also gives new discussion basis to a recent debate in the robotics industry. As general large models grow stronger, how much irreplaceable value remains in the robot brains embodied intelligence companies specially train. Last month, videos of GPT-6 Astra controlling robotic arms to draw the Golden Gate Bridge, grasp blocks, and imitate human operations went viral. A robotic arm priced at about $150, under Astra's control, could autonomously plan actions based on camera images, drawing the Golden Gate Bridge stroke by stroke. A more persuasive experiment came from RoboCurve. In its real-machine tests, Astra controlled dual robotic arms to place red blocks into a bowl, succeeding 19 out of 20 times. Such results led many to ask—since general large models already understand vision, space, reasoning, and can write control code, why do robotics companies still invest大量 time and money training their own VLA models from scratch. Some even proposed that the OpenAI of robotics might be OpenAI itself.
但身处具身智能行业的人，并不完全买账。9月10日的外滩大会上，自变量机器人创始人王潜谈到这种降维打击的说法时，判断目前可能性不大。他的理由是，模型的能力来自数据。编程模型变强，背后有大量代码数据和成熟的验证体系，机器人面对的物理世界，同样需要专门的数据、评测、硬件与基础设施，不会因为语言模型强大就自动解决。破壳机器人创始人许华哲也动手试过Astra。它能识别桌面物品，甚至抓起一根筷子，但一旦要求它用筷子进行更复杂的接触操作，表现就明显下降。在他看来，语义理解、空间理解的进步令人惊喜，复杂的物理交互却还是另一道题。
But people in the embodied intelligence industry do not fully buy it. At the September 10 Bund Conference, AgiBot founder Wang Qian, discussing this "dimensional reduction strike" claim, judged it currently unlikely. His reasoning: model capability comes from data. Programming models getting stronger is backed by large amounts of code data and mature validation systems; the physical world robots face also requires dedicated data, evaluation, hardware, and infrastructure, and will not be automatically solved because language models are powerful. Poke Robot founder Xu Huazhe also tried Astra hands-on. It can recognize desktop items, even pick up a chopstick, but once asked to use chopsticks for more complex contact operations, performance drops明显. In his view, advances in semantic understanding and spatial understanding are surprising, but complex physical interaction is another question entirely.
AA-Robotics恰好提供了一把新尺子。每款模型都获得一台Hiwonder NexArm机械臂的控制权。这是一台带五个关节和一个夹爪的六轴机械臂，桌面周围和机械臂手腕上共布置四个摄像头。模型起初只知道机械臂的名称、如何通过HTTP API与它通信，以及如何提交任务。它得自己看摄像头、理解物体位置、编写并运行程序，让机械臂完成动作。八项测试均为零样本尝试，不会为了当前题目再训练一次。首批五款模型的综合成绩已经公布，GPT-6 Astra得到69.4分，GPT-6.1 Sol为64.6分，Gemini 4 Argon为55.3分，Claude Opus 5.5和Claude Fable 5.1分别为48.8分、46.3分。不过，分数不是成功率。每道题满分100，没完全做成也可能拿到过程分。因此，Astra虽然总分最高，八道题中拿到满分的是四道，排第二的Sol，反而有五道满分。
AA-Robotics happens to provide a new ruler. Each model gets control of a Hiwonder NexArm robotic arm. This is a six-axis robotic arm with five joints and one gripper, with four cameras arranged around the desktop and on the arm's wrist. Models initially only know the arm's name, how to communicate with it via HTTP API, and how to submit tasks. It must look at cameras itself, understand object positions, write and run programs, and make the arm complete actions. All eight tests are zero-shot attempts, without retraining for the current problem. The first batch of five models' comprehensive scores have been announced—GPT-6 Astra scored 69.4, GPT-6.1 Sol 64.6, Gemini 4 Argon 55.3, Claude Opus 5.5 and Claude Fable 5.1 at 48.8 and 46.3 respectively. However, scores are not success rates. Each question is out of 100, and partial completion may earn process points. So although Astra has the highest total score, it got full marks on four of eight questions; second-place Sol actually got five full marks.
最简单的两项任务，五个模型全部过关，用夹爪按下红色按钮，以及把一只鳄鱼玩具完整放进指定方框。这些操作已经难不住今天的前沿大模型。再往后，差距迅速拉开。Astra能把瓶里的饮料倒进杯子，还能握着笔在纸上写出hello world，却在测量盒子长宽高的任务中拿了零分。Sol完成了收纳积木、形状匹配和写字，却没能倒好水。搭积木塔更是让所有模型都没拿到满分，Sol甚至一度搭出合格的塔，却因为后续操作前功尽弃。这一幕很能说明问题。在电脑里，模型发现答案后只需要停止生成，到了物理世界，它还必须判断眼前的物体是否已经放稳，任务是否真的完成，以及接下来的动作会不会把成果毁掉。
The two simplest tasks were passed by all five models—pressing a red button with the gripper and placing a crocodile toy completely into a designated box. These operations no longer challenge today's frontier large models. Further on, the gap rapidly widens. Astra can pour drink from a bottle into a cup and hold a pen to write "hello world" on paper, yet scored zero on the task of measuring a box's length, width, and height. Sol completed storing blocks, shape matching, and writing but failed to pour water well. Building block towers left all models without full marks; Sol even一度 built a qualified tower but ruined it with subsequent operations. This scene illustrates the problem well. In a computer, the model only needs to stop generating after finding the answer; in the physical world, it must also judge whether the object in front is stable, whether the task is truly complete, and whether the next action will destroy the achievement.
更惊人的反差，还在时间和成本上。五款模型中，除了Gemini 4 Argon平均耗时约52分钟，另外四款完成一项任务大约都要半小时。而Artificial Analysis给出的人工直接操作基线，平均只需约20秒。按红色按钮这样的小事，Astra花了2.2分钟，Claude Fable 5.1甚至花了约10分钟理解场景、规划动作，才开始第一次操作。钱也没少花。Astra平均每项任务的模型调用成本达到16.36美元，Sol只要1.93美元，分数却仅相差4.8分。对于未来要在家庭、仓库里连续工作的机器人来说，能完成一次，和能以合理成本每天完成上百次，是两回事。
The more astonishing contrast is in time and cost. Among the five models, except Gemini 4 Argon averaging about 52 minutes, the other four take about half an hour per task. Artificial Analysis's human direct operation baseline averages only about 20 seconds. For something as small as pressing a red button, Astra took 2.2 minutes; Claude Fable 5.1 even spent about 10 minutes understanding the scene and planning actions before starting its first operation. Money was not spared either. Astra's average model call cost per task reached $16.36; Sol only $1.93, with scores differing by just 4.8 points. For robots that will work continuously in homes and warehouses in the future, completing a task once and completing it hundreds of times daily at reasonable cost are two different things.
为什么只是移动一只机械臂，大模型就要琢磨这么久。Artificial Analysis拆解了整个任务过程。以Astra为例，它有约20%的时间用于理解环境和机械臂，约59%用于规划动作，实际机械臂运动只占约13%。Sol的情况类似，61%的时间在规划，机械臂运动也只有约13%。它们看起来像一个第一次走进实验室的人，先观察机械臂怎么动，再写程序、试探位置、检查画面，确认无误后才发出下一道命令。动作本身往往只需要几秒钟，前后的思考却可能持续数分钟。慢还不是全部问题。在六项涉及抓取的任务中，Astra的抓取成功比例为45%，Claude Fable 5.1为44%，而Sol只有28%。模型明明已经找到目标，夹爪却可能闭合在物体旁边，有时刚抓起来，瓶子又从指间滑落。
Why does just moving one robotic arm require large models to ponder so long. Artificial Analysis broke down the entire task process. Taking Astra as an example, about 20% of its time is for understanding the environment and robotic arm, about 59% for planning actions, with actual arm movement accounting for only about 13%. Sol's situation is similar—61% of time planning, arm movement also only about 13%. They look like someone entering a laboratory for the first time—first observing how the arm moves, then writing programs, probing positions, checking images, and only issuing the next command after confirming correctness. The action itself often takes only seconds, but the thinking before and after can last minutes. Slowness is not the whole problem. Among six grasping-related tasks, Astra's grasp success rate was 45%, Claude Fable 5.1 44%, and Sol only 28%. The model has clearly found the target, yet the gripper may close beside the object; sometimes just after picking up, the bottle slips from between the fingers.
这些错误很难靠再想一会儿全部解决。真实世界里的物体会滑动、倾倒、变形，手指和物体接触时还会产生摩擦与碰撞。机器人需要根据运动和接触反馈及时纠错，而不是每动一下都等待大模型重新规划。这也是专业具身模型仍然有发挥空间的地方。VLA、底层运动策略和控制器长期处理的，正是如何把视觉和任务目标快速转换成连续、稳定的动作。这份榜单还没有把通用Agent与专用VLA放到同一套条件下直接较量。八道桌面题目也无法代表全部机器人任务，更不用说双手协作、柔性物体操作和长时间自主工作。但它至少让争论前进了一步，通用大模型能完成真实操作，已经有了成体系的证据，它们距离高效、可靠地执行真实工作还有多远，也第一次在这组任务里被具体地量了出来。
These errors are hard to fully solve by "thinking a bit longer." Objects in the real world slip, topple, and deform, and fingers and objects generate friction and collision when in contact. Robots need to correct promptly based on motion and contact feedback, rather than waiting for the large model to re-plan after every movement. This is where specialized embodied models still have room to play. VLA, low-level motion policies, and controllers have long handled exactly how to quickly convert vision and task goals into continuous, stable actions. This leaderboard has not yet placed general Agents and specialized VLAs under the same conditions for direct competition. Eight desktop tasks cannot represent all robot tasks, let alone dual-arm collaboration, flexible object manipulation, and long-term autonomous work. But it at least advances the debate one step—general large models completing real operations now has systematic evidence, and how far they are from efficiently and reliably performing real work has been concretely measured in this task set for the first time.
AA-Robotics v0.1的首批名单中，还没有中国模型。那么，下一批会不会出现Qwen、Kimi、GLM或DeepSeek。如果放到同样的机械臂、同样的任务里，国产模型能否用更低成本追上甚至超过这些国际头部模型。这或许比单纯争夺一张榜单的名次更值得期待。机器人行业下一阶段要比的，已经不只是谁能让机械臂动起来，还有谁能让它动作快一点、抓得准一点，以及成功之后，知道什么时候该停手。
The first batch of AA-Robotics v0.1 does not yet include Chinese models. So will the next batch include Qwen, Kimi, GLM, or DeepSeek. If placed on the same robotic arm with the same tasks, can domestic models catch up with or even surpass these international leading models at lower cost. This may be more worth anticipating than simply competing for a leaderboard ranking. The robotics industry's next stage competition is no longer just about who can make the robotic arm move, but who can make it move faster, grasp more accurately, and—after succeeding—know when to stop.
