Imagine asking a robot to prepare breakfast for four people. Picking up a cup is one part of the job. The robot also needs to decide what to prepare, coordinate the kettle and toaster, notice that someone has moved a plate, and check that everyone has what they asked for. If breakfast is its responsibility every morning, it must remember changing preferences and schedules after today's meal is finished. Each request delegates a different kind of work.
设想让机器人为四个人准备早餐。拿起杯子只是其中一步。它还要决定准备什么、协调水壶和烤面包机、注意到有人挪走了盘子，并确认每个人都拿到了自己想要的东西。如果每天的早餐都归它负责，它还得在今天的早餐结束后，继续记住不断变化的偏好和日程。这些请求委托的是不同层次的工作。
Our roadmap paper, Envisioning Agentic Robotics: A New Era of Robotic Systems, starts from this shift in responsibility. It offers a functional definition of the robot's decision-making core, an evaluation framework, and a path from individual primitives to standing roles. These are proposals for organizing research and assessing progress. The breakfast scenario is an illustration of that path, not a demonstration of a system we have built.
我们的路线图论文《Envisioning Agentic Robotics: A New Era of Robotic Systems》从这种责任变化出发，给出机器人决策核心的功能定义、评估框架，以及从单个基础动作走向长期职责的研究路径。这些是组织研究、衡量进展的提议。早餐场景用来解释这条路径，并非我们已经搭建的系统演示。
Two decisions close the loop两类决策，形成闭环
Agentic robotics places an agent in charge of interpreting intent, organizing capabilities at run time, and revising its decisions from physical evidence. Learned action policies remain essential. They can already contain reasoning and feedback; the distinction concerns who takes responsibility for the overall goal, the next action, and the judgment that the work is complete.
在 agentic robotics 中，智能体负责理解意图、在运行时组织能力，并依据物理证据修订决策。学习得到的动作策略依然重要，也可以包含推理和反馈。这里的区别在于：由谁对整体目标、下一步行动以及工作是否完成的判断负责。
We call this functional role System 2, or S2. It must make two identifiable kinds of decision: how to proceed, and whether the outcome achieved its goal. Consider bringing a red cup to the dining table. Choosing a grasp is a decision about how to proceed. Checking that the cup actually reached the table, then deciding whether to finish or recover, closes the loop. A planner that never checks effects implements only part of the role. A detector that reports a success score supplies evidence; another component still has to decide what that evidence means for the task.
我们将这一功能角色称为 System 2，简称 S2。它必须作出两类可以从外部辨认的决策：接下来如何做，以及结果是否实现了目标。以把红杯子送到餐桌为例，选择抓取方式属于第一类决策；检查杯子是否真的到了桌上，再决定结束还是恢复，才形成闭环。只规划、不检查效果的组件只完成了部分职责。检测器给出的成功分数是证据，仍需要其他组件判断这些证据对任务意味着什么。
The definition does not require a particular network architecture or decision frequency. S1 supplies callable skills and modules; S0 maintains body control. One model may serve several roles, and several components may share one role. S2 can invoke a skill, write a program, or send a permitted action target directly to S0. Even in that last case, the foundation model does not replace the local real-time control loop.
这个定义不限定网络结构或决策频率。S1 提供可调用的技能与模块，S0 负责身体控制。同一个模型可以承担多个角色，一个角色也可以由多个组件共同完成。S2 可以调用技能、编写程序，或通过允许的接口直接向 S0 发送动作目标。即使采用最后一种方式，基础模型也不会取代本地实时控制循环。
The model needs a view of the world模型需要一份可信的世界状态
In our framework, Embodied S2 combines a foundation model with an embodied harness. The model provides knowledge and reasoning. The harness assembles its context, exposes tools, maintains plans and memory, and helps verify execution. Including the harness in S2 makes the system boundary explicit. A model can only decide from the observations it receives and act through the interfaces it is given.
在我们的框架中，Embodied S2 由基础模型与 embodied harness 组成。模型提供知识和推理；harness 则组织上下文、提供工具、维护计划与记忆，并协助验证执行结果。将 harness 纳入 S2，才能明确系统边界：模型只能根据收到的观测决策，也只能通过已有的接口行动。
For the breakfast robot, context includes where the cups were last seen, whether a person changed their request, and what the robot can safely reach. These facts have different sources and ages. A remembered location may help search, but a fresh observation may invalidate it. Persistent memory therefore needs revision as well as retrieval. It must also respect who may know or use a household member's preferences.
对准备早餐的机器人，上下文包括上次在哪里看到杯子、是否有人修改了请求，以及机器人能够安全够到哪里。这些信息的来源和时间不同。记住的位置有助于搜索，但新的观测也可能推翻它。因此，持久记忆不仅需要检索，还需要修订；对家庭成员偏好的使用，也应遵守知情和使用权限。
Interfaces need equally careful semantics. A tool accepting a command, a controller finishing its motion, the cup reaching the table, and the user's request being satisfied are four different events. Cancellation also needs an acknowledged outcome: requesting a stop does not establish that motion has stopped. These distinctions give the agent evidence it can reason about and give researchers a way to locate failures.
接口的语义同样需要精确。工具接受命令、控制器结束运动、杯子到达桌面，以及用户的请求得到满足，是四件不同的事。取消动作也需要确认结果：发出停止请求，不代表运动已经停止。区分这些状态，才能让智能体获得可用于推理的证据，也让研究者知道故障发生在哪里。
The world keeps moving while the model thinks模型思考时，世界仍在变化
Digital agents offer useful design principles: close the loop on evidence, define tool contracts, keep durable state, and reuse validated capabilities. Their most favorable setting, a coding sandbox, also offers readable state, executable tests, abundant data, and inexpensive resets. Physical robots must reconstruct state from partial observations, judge effects that may be hard to see, and act in environments that cannot always be reset.
数字智能体提供了值得借鉴的原则：依据证据形成闭环、明确工具契约、保存持久状态，以及复用经过验证的能力。在较为有利的编程沙箱中，状态可以读取，测试可以执行，数据丰富，重置成本也低。机器人则必须从局部观测重建状态，判断未必能直接看见的物理效果，并面对无法随时重置的环境。
Time changes the objective too. Suppose the robot observes a cup near the edge of a moving tray. A longer reasoning trace may improve its interpretation of the image while leaving it with an outdated target. We describe the resulting requirement as capped scaling: improve decision quality within the available physical deadline. This is a research objective, not an empirical scaling law established by our paper.
时间也会改变优化目标。假设机器人看到一个杯子位于移动托盘的边缘。更长的推理可能让它更准确地理解图像，却也让目标位置过时。我们用 capped scaling 描述由此产生的要求：在物理环境允许的时间内提高决策质量。这是研究目标，不是本文通过实验确立的缩放定律。
A complete system must remain responsive during deliberation. Local control continues, and a real-time event monitor can trigger predefined protective responses when an event cannot wait. S2 then decides what the change means for the task. Better reasoning, interruptible execution, and independent protections have to work together; a fluent explanation of a hazard does not establish safe behavior at the action interface.
完整系统在思考期间也必须保持响应。本地控制持续运行；遇到不能等待的事件，实时事件监控器可以触发预先定义的保护动作，随后由 S2 判断变化对任务意味着什么。更好的推理、可中断的执行和独立保护需要共同发挥作用。能够流畅解释风险，并不等于在动作接口上表现安全。
Evaluate the decisions and the system既评估模型，也评估完整系统
A benchmark score becomes more useful when we know what produced it. Our framework separates model-level probes from system-level evaluations. The former test capabilities such as knowledge and understanding under specified inputs. The latter test the loop that connects observations, decisions, tools, and physical outcomes. Comparing models requires controlling the harness; comparing harnesses requires controlling the model and other relevant conditions.
只有知道成绩由什么产生，基准分数才更有意义。我们的框架区分模型级测试和系统级评估：前者在明确输入条件下测试知识、理解等能力；后者评估连接观测、决策、工具与物理结果的闭环。比较模型时需要控制 harness，比较 harness 时则需要控制模型及其他相关条件。
We organize evaluation around knowledge, understanding, reasoning, tool use, memory, cross-embodiment, safety, efficiency, and cost. These dimensions overlap and play different roles. Cost measures resources rather than an additional ability. A profile across them can reveal a system that understands a scene well but verifies completion poorly, or one that succeeds only when deliberation is allowed to pause the world.
评估围绕知识、理解、推理、工具使用、记忆、跨本体能力、安全、效率和成本展开。这些维度彼此重叠，作用也不同；成本是资源指标，并非额外的一种能力。分别报告各维度，可以揭示系统是否善于理解场景却不善于判断完成，或是否只有在思考期间暂停世界才能成功。
The practical reporting requirements follow directly. A harness card should disclose the model and runtime versions, observations, budgets, time semantics, and number of runs. Live-world evaluation should accompany paused evaluation. Outcome checks should be independent enough to expose false completion claims, and reports should include human assistance. The paper draws on published results to motivate these requirements; it does not present a new experimental leaderboard or reproduce those studies.
由此可以得到具体的报告要求：harness 说明卡应披露模型和运行组件版本、观测、预算、时间语义及运行次数；暂停环境的评估应配套报告世界持续运行时的结果；结果核查应足够独立，能够暴露错误的完成声明；报告还应说明人工帮助。论文借助已有研究说明这些要求的必要性，并未建立新的实验排行榜，也未复现这些研究。
From a Primitive to a Standing Role从单个基础动作，到长期职责
The roadmap measures progress by what people can delegate. At L1, a robot executes a primitive such as picking up a cup and checks the result. At L2, it completes a task such as making and serving coffee, recovering when steps fail. At L3, it coordinates interdependent tasks, as in breakfast for four, managing timing, priorities, and shared resources. At L4, it sustains an agreed role across tasks and workflows, such as managing daily breakfast around the household's changing needs.
这条路线图以人们能够委托多少工作衡量进展。在 L1，机器人执行拿起杯子这样的基础动作，并检查结果。在 L2，它完成制作并送上咖啡这样的完整任务，在步骤失败后恢复。在 L3，它像准备四人早餐那样协调相互依赖的任务，管理时间、优先级和共享资源。在 L4，它跨任务、跨工作流程承担约定的职责，例如根据家庭不断变化的需求管理每日早餐。
Running longer does not by itself establish a higher level. Repeating a task for a day differs from deciding which tasks a standing responsibility requires. Spatial scope is another axis: work can expand from a workcell to a room, building, or site without changing its level. Every claim should therefore state the responsibility, duration, spatial scope, complexity, and human assistance behind it.
运行更久，本身并不能证明等级更高。重复一天同样的任务，与自主决定履行长期职责需要完成哪些任务，是不同的能力。空间范围是另一条轴：工作可以从工作单元扩展到房间、建筑或园区，而责任等级不变。因此，每项能力声明都应说明背后的责任范围、时长、空间、复杂度和人工帮助。
The evidence reviewed in the paper supports progress on primitives and tasks, with workflow-level evidence beginning to appear in company demonstrations. It does not establish sustained roles in unstructured environments without human repair. Closing that gap calls for reasoning within physical constraints, persistent and verifiable control, efficient adaptation to new environments and bodies, and continual improvement from physical experience.
论文回顾的证据支持基础动作和任务层面的进展，公司演示中也开始出现工作流程层面的证据。但这些证据尚未证明机器人能在非结构化环境中，无需人工补救地持续承担长期职责。要缩小这一差距，需要研究物理约束下的推理、持久且可验证的控制、对新环境和新本体的高效适应，以及从物理经验中持续改进。
The intended outcome is practical: useful work that meets agreed criteria while reducing the work left to people. Progress toward it should be visible in verified deliveries, reliable recovery, and disclosed intervention burden. That is the standard our roadmap asks the field to make measurable as robots take on broader responsibility.
我们希望实现的结果很具体：机器人按照约定标准完成有用的工作，同时减少留给人的工作。经过核实的交付、可靠的故障恢复，以及公开的人工介入负担，应当让进展清晰可见。这正是路线图希望在机器人承担更广责任时，推动研究领域共同衡量的标准。