On May 13, 2026, Li Xiang, CEO of Li Auto, proposed a highly interesting perspective during a public dialogue: autonomous driving is the first half of embodied AI, and general-purpose humanoid robots are the second half. This judgment quickly spread across the industry, with many comparing it to Wang Chuanfu's earlier assertion that electrification is the first half and intelligence is the second half. Focusing solely on the claim that autonomous driving is the first half of embodied AI, is this statement accurate?
01. Technical Foundation or Talent Migration?
In fact, many people support this view, believing that autonomous driving has already accumulated considerable depth in core technologies such as perception, decision-making, and control, and these capabilities can be directly transferred to broader embodied AI scenarios. From a system architecture perspective, autonomous driving and embodied AI indeed share similar frameworks; both rely on sensor data to build environmental models in real time and make decisions. Technical concepts like end-to-end large models, VLA (Vision-Language-Action model), and world models are also being widely discussed and advanced in both fields simultaneously.
However, others argue that the "first half" narrative creates a logical inertia, conflating chronological sequence with causal inevitability. The core problem autonomous driving needs to solve is relatively singular: safe navigational movement, which hinges on environmental perception, path planning, and collision avoidance. In contrast, embodied AI must simultaneously address three major challenges: navigational movement, motion control, and dexterous manipulation. Navigational movement is merely the entry ticket; the latter two are the true core challenges.
Technically, the differences between the two are quite obvious. Autonomous driving pursues efficient and safe movement under highly structured traffic rules, whereas embodied AI must cope with broader and more complex unstructured environments. To make a vivid comparison, autonomous driving solves the problem of keeping a car from colliding with objects on a 2D plane, while embodied AI handles making devices interact with objects in 3D space. The physical constraints and operational targets they face are entirely different.
Given these significant technical differences, what is the true foundation for the "first half" argument? Zhijia Qianyan believes that the key to the first half might lie not in the technology itself, but in the people. At this stage, engineers who have emerged from the autonomous driving track are collectively flocking into the embodied AI field, carrying capabilities tempered by mass-production engineering. Regarding this trend, Zhijia Qianyan has also analyzed: As autonomous driving talents flock to embodied AI, is it an overwhelming advantage or will they face adaptation difficulties?
According to statistics, Horizon Robotics alone has seen at least 14 core technical and management personnel leave to start businesses, 13 of whom have entered the embodied AI track; former Li Auto CTO Wang Kai, former head of autonomous driving technology R&D Jia Peng, and others have also successively joined the embodied AI entrepreneurship wave. The experience of these individuals in the autonomous driving field has equipped them with a complete set of methodologies for autonomous driving development. They know how to define complex problems, how to break down tasks under uncertainty, and how to make critical trade-offs under resource constraints.
02. Can VLA Survive Outside the Car?
In 2026, the VLA model has become the most watched technological link connecting autonomous driving and embodied AI. Building upon the vision-language model, VLA introduces action tokens, enabling the model to jointly learn from paired visual, linguistic, and robot trajectory data. Li Auto released MindVLA-o1 at GTC 2026, stating that autonomous driving is merely the starting point for physical AI, and such foundation models will drive new embodied AI paradigms in the future; Xiaomi's open-sourced Xiaomi OneVL also explicitly unifies VLA, world models, and latent space reasoning into a single framework, positioning it as a foundational component for both autonomous driving and embodied AI.
VLA technology can be shared, but there is an unavoidable issue: latency. Currently, the inference latency of VLA models is generally between 100 and 300 milliseconds. For autonomous driving, this latency has a significant impact. When the vehicle speed is 72 km/h, the car will travel 6 to 8 meters forward during the model inference period. In urban intersections or highway car-following scenarios, this is enough to cause collisions or lane departure accidents. Therefore, autonomous driving requires end-to-end latency to be compressed to within 50 milliseconds.
However, for humanoid robots, most manipulation tasks involve grasping stationary workpieces, screwing bolts, etc., which have a higher tolerance for latency. Even if inference takes 200 milliseconds, the physical movement of the robotic arm itself will take hundreds of milliseconds to several seconds. The model output can be issued as setpoints; latency only affects the overall cycle time and will not directly lead to task failure. Latency only becomes a critical consideration when dealing with rapid interactions, but such scenarios are not common in the embodied AI field. Thus, the same model architecture has vastly different real-time requirements on different physical carriers.
World models actually face a similar situation. In the first half of 2026, Geely released the WAM (World Action Model), XPeng released X-World, and Xiaomi launched the Xiaomi Auto World Model. Almost every leading autonomous driving company is deploying world models. World models are regarded as a key puzzle piece for achieving Artificial General Intelligence (AGI) and are also seen by the industry as a critical technology to break through the generalization bottleneck of embodied AI. However, the tasks undertaken by world models in the two scenarios of autonomous driving and embodied AI are not the same.
On the vehicle side, it mainly performs scenario deduction and planning verification to test and optimize autonomous driving strategies; whereas in the embodied AI field, it needs to achieve true physical-level interactive reasoning, such as the mechanical modeling involved behind actions like grasping and assembly. In the two fields, world models focus on trajectory-level prediction on one hand, and handle physical contact and manipulation planning on the other. The depth of understanding required for physical laws differs by more than one level between the two.
03. From Moving to Working?
Industry data also reveals the development potential of embodied AI. In 2025, the market size of China's embodied AI was approximately 915 billion RMB, and it is expected to reach 1,090.4 billion RMB in 2026. The number of related enterprises nationwide has exceeded 10,000. The 2026 "Report on the Work of the Government" listed embodied AI as a key frontier technology for cultivation. The embodied AI special action jointly launched by the Ministry of Industry and Information Technology (MIIT) and the State-owned Assets Supervision and Administration Commission (SASAC) also quantified the landing target to the scale of 10,000 units for the first time. It can be seen that in the field of embodied AI, enthusiasm from both the policy and market sides is fully in place.
The market potential is promising, but has the actual technological maturity reached the standard? Industry insiders judge that embodied AI is currently still in the transition stage from L1 to L2 and L3. Referring to the group standard "Classification of Intelligence for Humanoid Robots" (T/CIE 298-2025) released in 2025, L1 is basic perception, L2 is autonomous decision-making, and L3 is complex tasks. Humanoid robots can be deployed in some specialized scenarios, but there is still a significant distance to true cross-scenario generalization. Liu Dong, founder of Xingyuanzhi, analogized the current deployment status of embodied AI to autonomous driving in 2015 and 2016, when L2 advanced driver-assistance systems (ADAS) had just begun to be deployed at scale.
Embodied AI has indeed made some progress at this stage, but to say it will replace human labor on a large scale like autonomous driving is still a long way off. Looking back at the journey of autonomous driving, from the widespread expectation in 2017 that L4 mass production could be achieved in three to five years, to 2026 when the L2 penetration rate has just exceeded 70% and L3 has just begun commercial pilot programs, it has taken nearly a decade. The scenarios embodied AI must face are much more complex than driving, and physical interactions are far richer. This cycle will most likely be longer, not shorter.
04. Is Autonomous Driving the First Half of Embodied AI?
Returning to the initial question, does the assertion that autonomous driving is the first half of embodied AI hold true?
If we understand the "first half" as occurring first in time and being validated first commercially, then it basically holds true. Autonomous driving is indeed the most industrialized application scenario under the concept of embodied AI. It was the first to complete the data closed-loop from algorithms to mass production, the first to establish the supply chain system for sensors, chips, and operating systems, and the first to cultivate a group of engineering talents who know how to deploy AI models into the physical world.
However, if we understand the "first half" as a necessary technical path or sufficient preparation in capabilities, then it does not strictly hold true. The perception and decision-making capabilities accumulated by autonomous driving can be partially reused, but motion control and dexterous manipulation are new fields. An algorithm stack designed for four-wheeled cars cannot directly drive a bipedal robot to perform grasping and assembly. More critically, the physical laws, interaction targets, and costs of failure faced by the two are completely different. A loss of control in a car may cause life-threatening dangers, while a failed robot grasp might just require a retry. This difference also determines that the technical routes, safety standards, and engineering methodologies of the two will head in different directions.
It can be said that the "first half" is a useful metaphor. It helps the industry understand the connection and evolutionary sequence between the two fields, but a metaphor is just a metaphor after all; it simplifies reality and obscures differences. Autonomous driving is not an inevitable prelude to embodied AI, just as electrification is not an insurmountable prerequisite for intelligence. What truly connects the two fields are the talents tempered in the autonomous driving field and the methodologies they carry.
From a more macro perspective, the market for physical AI is forming at this stage, and various car companies are successively releasing physical AI solutions. Zhijia Qianyan has also noticed this phenomenon: What are the physical AI large models successively released by various car companies, and what are their advantages? Strategy& estimates that the global physical AI market size will reach 430 billion euros by 2030, with autonomous driving and embodied AI as the core applications. In this larger technological landscape, autonomous driving and embodied AI are more like two parallel tracks, sharing some underlying technologies but also facing unique engineering challenges respectively. Undeniably, there is indeed a flow of talents, data, and methodologies between them, but this flow is bidirectional and gradual, rather than unidirectional and decisive.
#AutonomousDriving #EmbodiedAI