Zheng Nanning: Embodied AI Is Not Humanoid Robots, Evaluation Isn't About Resembling Humans
Zheng Nanning, an academician of the Chinese Academy of Engineering and director of the Institute of AI and Robotics at Xi'an Jiaotong University, published an article on embodied intelligence growing in the physical world.
He points out that a major direction for artificial intelligence is to enable machines not only to process information and reason about complex problems, but also to flexibly complete various tasks in real, changing environments.
From fine operations on industrial production lines to autonomous walking in complex terrain to dynamic movements in performance settings, robots are appearing frequently across scenarios.
Behind these capabilities, AI is undergoing a significant shift: from being able to compute to being able to act, and further evolving into an intelligent form that continuously interacts, learns and evolves in the physical world.
This shift signals a broader transformation in how intelligence is understood and deployed.
It must be clarified that embodied intelligence is not equal to humanoid robots; humanoid form is only one carrier. The key lies in whether the intelligent system can truly enter a real environment, form a closed loop of perception, cognition, decision, action and feedback, and continuously adjust its behavior within this loop, improving internal models and strategies through learning.
Therefore, the criterion for evaluating embodied intelligence is not whether its form resembles a human, but whether it can truly solve real-world problems.
An intuitive way to understand embodied intelligence is to return to how an infant comes to know the world. Infants do not first learn concepts or language; they build understanding through constant touching, grasping, falling and retrying. Infant cognition is not detached from the body but gradually formed through body-environment interaction: vision provides spatial structure, touch provides physical feedback, and motor experience continuously corrects predictions about the world.
Intelligence is not merely a computational process existing in the brain; it is a product of body-environment interaction. Thus, the body is not just a tool for executing commands but a component of the cognitive system, participating in perception, understanding and action. A typical embodied intelligence system consists of three parts: a perception system for acquiring environmental information, a decision system for forming understanding and planning, and an execution system for acting on the physical world.
For machines to truly enter the world and act in it, embodied intelligence requires multiple capabilities. First, world model formation: the system needs to build an internal simulatable environment so it can predict and reason before acting.
Second, multimodal representation: real-world information is highly heterogeneous, including vision, language, touch and sound; embodied intelligence must integrate these into a unified semantic space.
Third, causal reasoning: current large models mainly learn based on statistical correlations and have limited grasp of causal structures; their judgments may sometimes seem linguistically reasonable but do not hold in the physical world.
Fourth, integrated generation and action: an end-to-end system that jointly models vision, language and action, enabling robots to generate action strategies from natural language instructions and to execute them through control systems and actuators.
The development of embodied intelligence relies on two fundamental conditions: high-quality multimodal data as fuel, and high-fidelity simulation environments as training grounds. Training in simulation and then transferring to the real world has become an important technical path, but how to narrow the sim-to-real gap remains an urgent problem.
Current systems still have limitations: the black-box problem, where model decision-making is insufficiently interpretable; the contradiction between computing power and deployment, as large models are difficult to run in real time on edge devices; and an immature industrial ecosystem, with standards and interface specifications still evolving.
In industrial manufacturing, intelligent systems are moving from execution tools to intent-driven systems. In healthcare and rehabilitation, they are becoming important assistive tools for doctors. In urban governance and public safety, drones and ground robots collaborate on infrastructure inspection tasks. In transportation, autonomous driving systems are among the embodied intelligence forms closest to large-scale application.
The essence of embodied intelligence is not to make machines resemble humans, but to let intelligence truly enter the real world, forming continuous learning and adaptation capabilities in complex, uncertain environments.