机器人还没有身体
Robots don't have bodies. Yet.
2026 年的人形机器人有作动器、传感器，越来越多还运行着一颗在板载 GPU 上跑的基础模型作为”大脑”。它能走路、能操作物体，能回答关于自己刚才做了什么的问题。
但它仍然没有身体——不是在具身认知所说的那种意义上的身体。
“拥有身体”真正要求的是什么
借用 Andy Clark、Linda Smith 和更早的实施认知主义者的观点：
- 通过世界闭合的感觉-运动耦合。 身体不是大脑指挥的外设；身体是大脑预测进入的底质。预测、误差、行动构成一个单一的回路。
- 可供性地景。 世界是按身体能对它做什么来被感知的。一段楼梯之所以”可攀登”，只在腿长和扭矩预算的相对意义上成立。
- 稳态（homeostasis）。 身体有筹码——它会损耗、会疼、会饿。没有这些，“自我保存”就只是一条指令，而不是一种驱力。
当下的人形机器人对 #1 的满足是弱的（回路存在，但被 LLM 的文本世界模型主导）；对 #2 是脆弱的、由数据学到的方式；而对 #3 完全没有满足。电池百分比是最接近的类比，但它对能动性的构成性远不如疼痛之于我们。
真正有趣的前沿
有趣的问题不是要不要把模型做得更大。是我们能否做出这样的机器人：它的策略网络以自身的稳态状态作为主要损失函数，任务奖励只是次要的。这是一个非常不同的优化曲面，可能会产出在质上不同的行为：会累的机器人，会偏好某条路线（出于能量原因）的机器人，会回避伤害（不是因为被指令所要，而是因为伤害真的疼）的机器人。
我会论证：那才是”机器人”不再是”自动机器”的隐喻、而开始更接近”主体”的时刻。
A humanoid robot in 2026 has actuators, sensors, and increasingly a foundation-model “brain” running on a beefy onboard GPU. It can walk, manipulate, and answer questions about what it just did.
It still does not have a body, in the sense embodied cognition means it.
What “having a body” really demands
Drawing from Andy Clark, Linda Smith, and the older enactivists:
- Sensory-motor coupling that closes through the world. The body isn’t a peripheral that the brain commands; it is the substrate that the brain predicts into. Predictions, errors, and actions form a single loop.
- Affordance landscapes. The world is perceived in terms of what the body can do with it. A staircase is “climbable” only relative to leg length and torque budget.
- Homeostasis. Bodies have stakes — they degrade, hurt, hunger. Without that, “self-preservation” is an instruction, not a drive.
Current humanoids satisfy #1 weakly (the loop exists, but is dominated by the LLM’s textual world-model), #2 in a brittle, learned-from-data way, and #3 not at all. Battery percent is the closest analogue, and it is not constitutive of agency the way pain is for us.
The interesting frontier
The interesting question isn’t whether to scale models bigger. It’s whether we can make robots whose policy networks are trained against their own homeostatic state as a primary loss, with task rewards as secondary. That is a very different optimisation surface, and might produce qualitatively different behaviour: robots that get tired, that prefer one route over another for energetic reasons, that avoid damage not because instructed to but because damage hurts.
That, I’d argue, is when “robot” stops being a metaphor for “automated machine” and starts being something closer to “agent”.
Related: Predictive processing 101, The loop and the self.
💬展开评论 / Show comments