EmbodiedLab：
评测为了知道而行动的具身智能体EmbodiedLab: Benchmarking Embodied Agents That Act to Know

30 个力学、电学、光学、热学和化学的教科书级实验，都在同一间仿真实验室里，由一条机械臂动手完成。悬停暂停，点击观看完整录像。30 textbook experiments in mechanics, electricity, optics, thermal physics and chemistry, all in one simulated laboratory and all carried out by one robot arm. Hover to pause; click to watch the full run.

每段是一次完整示范运行的精选片段，按步骤加速剪辑。大画面是最能看清操作的相机，右侧是另外两台相机。Each clip is a highlight edit of one complete demonstration run, sped up step by step. The large view is the camera that shows the work best; the other two cameras are on the right.

传统机器人执行的是事先为任务设计好的流程；具身智能体拿到的是一个目标，要自己决定和什么交互、怎么交互、什么时候交互。现实里的很多目标无法朝着已知的终态直接去做：智能体得先改变系统的物理状态，观察反馈，不断探索，直到达成目标，或者弄清目标到底需要什么。我们把这叫作“为了知道而行动”：行动是为了获取信息，改变物理状态是手段。由前沿大模型配上工具和技能组成的具身智能体，已经开始承担这类任务。A traditional robot carries out a procedure engineered in advance for its task; an embodied agent is given a goal and must decide for itself what to interact with, how and when. Many real goals cannot be reached by acting toward a known target: the agent must first change the physical state of the system, observe the feedback and explore until it reaches the goal or learns what the goal requires. We call this acting to know: action whose purpose is to uncover information, with changing the physical state as its means. Embodied agents that couple frontier models with tools and skills have begun to take on tasks of this form.

现有基准把这个过程的两端分开来测。操作类基准给出场景应达到的状态，看机器人能不能做到；目标是给定的，不需要弄清任何事。而且只检查最终状态，看不出这个状态是怎么达到的，智能体可以绕开任务本来要考的动作。科学发现类基准确实要求智能体弄清某个规律或参数，但实验只是一条文字指令，结果以文字或精确数值返回，智能体和信息之间没有任何物理动作。真实任务里两者是耦合的：怎么动手决定了能观察到什么，观察到什么又决定了下一步怎么动手。Existing benchmarks evaluate the two ends of this process apart. Manipulation benchmarks name the state the scene should reach and check whether the robot reaches it; the target is given, so nothing has to be found out. A final-state check also cannot tell how the state was reached, so an agent may reach it by a route that skips the behavior the task is meant to test. Scientific-discovery benchmarks do ask the agent to find out a hidden law or parameter, but an experiment there is a text command and its result returns as text or as an exact value, so no physical action stands between the agent and the information. In a real task the two are coupled: how the agent acts decides what it can observe, and what it observes decides how it acts next.

科学定量实验是进入这一场景的一个可控入口。实验给出的是目标而不是步骤：测量一个性质（比如透镜焦距），预测一个响应（比如抛出的小球落在哪里），或者把装置调到满足某个指标。涉及的物理定律都很熟悉，但每个实例的参数都重新抽取，结果只能在眼前这套系统上测出来；又因为结果是物理量或物理状态，可以对智能体隐藏，并精确打分。难度来自物理本身和实验方法，就像教学实验室一样，而不是人为设置的障碍。Quantitative science experiments are a controlled entry point to this setting. An experiment poses an objective rather than a procedure: measure a property, such as a focal length; predict a response, such as where a launched ball will land; or configure the apparatus to meet a specification. The laws involved are familiar, yet the parameters are drawn anew for every instance, so the result must be established on the system at hand; because it is a physical quantity or state, it can be hidden from the agent and scored exactly. The difficulty comes from the physics and the experimental method, as in a teaching laboratory, rather than from obstacles added on purpose.

任务、环境和评分都由“为了知道而行动”推出。任务说明只给目标、器材和规则，不规定方法。实验室和真实实验室一样是开放的：台面上的仪器、工具和材料比任何一种方法需要的都多，每一步操作都是物理的，读数出现在刻度、指针和屏幕上，由智能体通过三台相机自己去读，或者来自数据记录仪的记录。一次运行只有同时满足三点才算成功：结果落在由仪器精度确定的容差内，所依据的读数是这次运行里记录下来的，实验台恢复到任务要求的状态。所以评分看的是结果怎么得来的，而不只是最后留下的状态：即使结果和目标吻合，只要不是基于这次运行里、在要求条件下记录的读数，也不算成功。每个实验都验证过能通过公开接口在容差内完成。Task, environment and metric each follow from acting to know. The task card states the objective, the apparatus and the rules, and prescribes no method. The laboratory is open, as a real one is: the bench offers more instruments, tools and materials than any one method needs, every operation is physical, and readings appear on scales, pointers and displays that the agent reads through its cameras, or in the records of data-logging instruments. A run succeeds only if its result lies within a tolerance set by the instruments, rests on readings recorded in that run, and leaves the bench in the state the task requires. A run is thus scored on how its result was obtained, not only on the state it leaves: a result that matches the target fails if it does not rest on readings recorded in the run under the required conditions. Every experiment is verified to be solvable within its tolerance through the public interface.

30 个实验分布在 14 个器材平台上：力学 10 个、电学 6 个、光学 5 个、热学 3 个、化学 6 个，都是教科书实验或它的直接延伸。按层级分，18 个是基础实验，事先固定的标准方案就能达到容差；12 个是进阶实验，要自己选测量条件、根据已有读数决定下一步、辨识模型，或者按自己得出的结果再动手。每个实例都会重新抽取隐藏参数，答案背不下来。每张卡片还列出这个实验要求的关键证据：所有读数都必须引用这次运行自己的观测或记录，并且和仪器当时的显示一致，卡片上写的是在此之外还要满足的条件。The 30 experiments run on 14 apparatus platforms: 10 in mechanics, 6 in electricity, 5 in optics, 3 in thermal physics and 6 in chemistry, all textbook experiments or direct extensions of them. By tier, 18 are basic, where a standard design fixed in advance meets the tolerance, and 12 are advanced, where the agent must choose the conditions, decide each next condition from the readings so far, identify a model or act on its result. Every instance draws fresh hidden parameters, so answers cannot be memorised. Each card also gives the evidence the experiment requires: every reading must be cited from the run's own observations or records and match what the instrument showed, and the card gives what must hold beyond that.

每个模型都在其开发方发布的编程智能体里运行，通过一个小巧的 Python 客户端操作实验室，并可以在上面自己写代码，搭建感知、控制和数据分析的工具。评分只依据实验过程的记录，不看模型自己怎么说。Each model runs in the coding agent its developer ships for it and acts through a small Python client, on which it builds its own tools for perception, control and analysis. Scoring uses only the records of the run, never what the model claims.

因为前沿模型都熟悉这五个学科的教科书知识，我们不按学科，而是按两条在任何智能体运行之前就定好的轴来描述实验：层级说明标准方案够不够用，能力维度说明交互难在哪里。另外，接口分成几档，逐档把更多的动手部分交给工具，在同一批实例上比较，就能看出每一档改变了什么。Since frontier models know the textbook science of all five disciplines, we describe the experiments instead by two axes fixed before any agent ran: the tier says whether a standard design suffices, and the capability dimension says where the interaction is hard. Separately, the interface comes at nested levels that hand more of the acting to tools; comparing them on the same instances shows what each level changes.

我们报告成功率和过程分。一次运行要结果在容差内、所依据的读数是本次记录的、实验台恢复到要求的状态，才算成功。复原实验台是实验的一部分，有两个理由。在为了知道而行动中，智能体必须在很长的过程里记住自己做过的每一处改动，除了目标要求保留的，其余都要还原；而且实验结束时装置应处于安全、可直接再用的状态，比如关掉加热器、断开电源、零件放回原位。过程分记录失败的运行走到了哪一步。每个实验先对它的各次运行取平均，再对实验取平均，所以每个实验的分量相同。排名按 T0 成功率。除特别说明外，下面的结果都在 T0 档：智能体只拿到相机画面、机械臂状态、关节目标和夹爪指令，其余都靠自己写代码完成。We report the success rate and a process score. A run succeeds only with a result within tolerance, resting on readings recorded in the run, on a restored bench. Restoring the bench is part of the experiment for two reasons. In acting to know the agent must keep track of every change it made over a long episode and undo all but those the objective requires; and a laboratory experiment ends with its apparatus in a safe state, ready for the next use, such as the heater off, the circuit unpowered and every part back in place. The process score records how far a failed run got. Each experiment is averaged over its runs first and then over the experiments, so every experiment weighs the same. Agents are ranked by the T0 success rate. Unless stated otherwise, results are at T0, where the agent gets only the camera images, the robot state, joint targets and gripper commands, and writes everything else itself.

• 复原实验台拉开了最强三个模型的差距。GPT-6 Astra、GPT-6.1 Sol 和 Claude Opus 5.5 把结果测进容差的比例接近（51.7% 到 55.0%），成功率的差距主要来自复原：GPT-6.1 Sol 有 23% 的运行结果准确却没把实验台复原，比其他模型都多。Restoring the bench separates the three strongest agents. GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 know the result within tolerance at similar rates (51.7% to 55.0%), so their success rates differ mainly in restoring: GPT-6.1 Sol leaves the bench unrestored after an accurate result in 23% of its runs, more than any other agent.

每一档改变了什么What each level changes

T1 提供感知和无碰撞的手部运动，T2 再加上操作技能，每一档都把更多的动手部分交给工具，而弄清答案始终是智能体自己的事。T1 supplies perception and collision-free hand motion and T2 adds manipulation skills, so each level takes over more of the acting, while finding out stays with the agent.

• 感知和运动规划帮助不大。从 T0 到 T1，成功率只提高 1.7 到 10 个百分点；提高的地方在观察上：GPT-6 Astra 的精细读数从 42.9% 升到 71.4%，Claude Opus 5.5 的长程多阶段也从 42.9% 升到 71.4%。规划好的运动能避开碰撞，但在哪里夹、怎么放下，仍要智能体自己决定。Perception and motion planning change little. From T0 to T1 the success rate rises by 1.7 to 10 points, and where it rises, it rises through observing: from 42.9% to 71.4% in fine reading for GPT-6 Astra and in long-horizon work for Claude Opus 5.5. Planned motion avoids collisions but leaves to the agent where to grasp a part and how to set it down.
• 操作技能帮助更大。在 T2，成功率提高 10 到 27 个百分点：DeepSeek 从 13.3% 到 40.0%，达到 GPT-6 Astra 在 T0 的水平；Kimi 从 20.0% 到 33.3%；GLM 从 3.3% 到 13.3%。提高最多的是精细操作与接触，以及多点采集与数据处理。Manipulation skills change more. At T2 the success rate rises by 10 to 27 points: DeepSeek from 13.3% to 40.0%, the T0 level of GPT-6 Astra; Kimi from 20.0% to 33.3%; GLM from 3.3% to 13.3%. The gains are largest in fine manipulation and in multi-point acquisition and analysis.
• 技能补不上的是实验本身。这三个模型在 T2 只有 47% 到 77% 的运行把读数采够，Claude Opus 5.5 在 T1 是 90%；在长程多阶段实验上，它们的成功率最多只有 14.3%；没有一个追上 Claude Opus 5.5 在 T1 的 53.3%。What the skills do not supply is the experiment itself. At T2 the three agents complete their readings in only 47% to 77% of the runs, against 90% for Claude Opus 5.5 at T1; they reach at most 14.3% in long-horizon work, and none reaches the 53.3% of Claude Opus 5.5 at T1.
• 每个实验都有模型做成过。30 个实验都在某一档被至少一个模型做成；T0 没有模型做成的三个，黑箱电阻、透镜焦距和双透镜成像，Claude Opus 5.5 在 T1 全部做成。Every experiment is solved by some agent at some level. The three that no agent solves at T0, the black-box resistor, the focal length of a lens and the two-lens relay, are all solved by Claude Opus 5.5 at T1.

所有模型使用相同的实验实例和任务说明。每个智能体都使用其编程智能体的中间推理档：GPT-6 Astra、GPT-6.1 Sol 和 Claude Opus 5.5 用 medium，这是 Codex 和 Claude Code 的默认档；DeepSeek-V4.1-Flash-1017、Kimi K3 和 GLM-5.3-Flash 开启思考、用 high，这是它们编程智能体的中间档；除推理档位外，都使用默认设置。All models get the same experiment instances and task cards. Every agent runs at the middle reasoning setting of its coding agent: GPT-6 Astra, GPT-6.1 Sol and Claude Opus 5.5 at medium effort, the default of Codex and Claude Code, and DeepSeek-V4.1-Flash-1017, Kimi K3 and GLM-5.3-Flash with thinking enabled at high effort, the middle setting of theirs; apart from the reasoning setting, every agent runs with the defaults of its harness.

每个实验都有一段示范录像，可以从三台相机同步观看。Every experiment has a demonstration run, which can be watched from the three synchronised cameras.

每次失败的运行都归到它第一个决定性的错误：智能体没有纠正、之后再也不可能成功的最早那个错误。原因按“为了知道而行动”的四个阶段分组：对器材动手、观察器材、知道该做什么和何时停下、恢复实验台。下面是 T0 档的结果。Each failed run is attributed to its first decisive error: the earliest error the agent does not correct and after which success is no longer possible. The causes follow the stages of acting to know: acting on the apparatus, observing it, knowing what to do and when to stop, and restoring the bench. Results are at T0.

• 动手是每个模型最大的失败原因。动手、观察和复原这些与装置的交互合计占各模型失败的 75% 到 100%，“知道”最多占 25%。DeepSeek、Kimi 和 GLM 的失败里，64% 到 69% 出在动手上，单是接触就占 45% 到 62%：夹取打滑、零件倾倒、按键按偏。它们很少败在“知道”上（不超过 8%），因为很少走到那一步。Acting is the largest cause for every agent. Acting, observing and restoring together, all interaction with the apparatus, account for 75% to 100% of every agent's failed runs, knowing for at most 25%. For DeepSeek, Kimi and GLM it accounts for 64% to 69% of the failed runs, and contact alone for 45% to 62%: grasps that slip, parts that tip over, keys that are missed. Their runs seldom fail at knowing (8% or less), because few of them get that far.
• 最强的三个模型，失败分散到后面几个阶段。动手降到 35% 到 42%，“知道”占 15% 到 25%，结果准确却没收拾好实验台占 25% 到 33%。读错数的情况不多，但一旦发生就是致命的：它占 Claude Opus 5.5 失败的 23%，是第二大原因。For the three strongest agents the failures spread over the later stages. Acting falls to 35% to 42%, knowing accounts for 15% to 25%, and an untidy bench after an accurate result for 25% to 33%. Observing errors are rarer but final: they account for 23% of the failed runs of Claude Opus 5.5, its second largest cause.

决定性的错误通常很小：唯一的样品掉了，一位数字读错，滑块读到了导轨刻度的另一条线上，滑块留在了导轨上。The decisive error is usually small: the only sample dropped, one digit misread, a rider read off the wrong mark of the rail scale, a glider left on the track.

每张图都截自这一局智能体自己拿到的相机画面；点开可以看出错前后的录像，左、右、腕部三个视角同步播放。Each image is a camera frame the agent itself received in that run; open it to watch the moments around the error from the left, right and wrist cameras at once.

如果 EmbodiedLab 对你的研究有帮助，请引用我们的论文。If you find EmbodiedLab useful in your research, please cite our paper.

@article{gao2026embodiedlab,
  title   = {{EmbodiedLab}: Benchmarking Embodied Agents That Act to Know},
  author  = {Gao, Hongcheng and Zheng, Zelin and Zhou, Alex and Zhu, Jay and Ge, Shijia and Wang, Yazhe and Li, Xiping and Zeng, Jianshu and Wu, Di and Shangguan, Xuan and Wan, Yexing and He, Lingyu and Wang, Yongli and Yu, Xiaohan and Jia, Zhiqi and Liu, Yue and Hooi, Bryan and Dong, Yinpeng and Zhang, Wentao and Wu, Sihang and He, Xiao},
  journal = {arXiv preprint arXiv:2610.35385},
  year    = {2026}
}