In recent years, autonomous driving companies have been focused on solving how machines understand road environments and make driving decisions based on changes among surrounding traffic participants. Now, these companies are attempting to extend their technological accumulation into areas such as humanoid robots and embodied AI.
In August 2026, XPeng announced that its robotics business had completed its first round of financing, raising over USD 900 million with a post-money valuation exceeding USD 6.3 billion. XPeng disclosed that its IRON humanoid robot adopts a fully self-developed full-stack system, covering the robot body, chips, controllers, motion modules, and dexterous hands.
Prior to this, Chery also showcased its embodied AI products at the Global Innovation Conference, stating that its robotics business reuses approximately 70% of its R&D technology.
Beyond automakers, autonomous driving companies are accelerating their layout in embodied AI. ZhiXing Technology is entering the robotics sector through its subsidiary EmoStar Robot, while Desay SV Automotive has launched an AI computing terminal for the robotics field, attempting to extend its intelligent driving technical architecture to new product forms.
As autonomous driving companies begin building robots, can autonomous driving technology truly support embodied AI?
01 Why Are Autonomous Driving Companies Starting to Focus on Robots?
The entry of the autonomous driving supply chain into the robotics field is not accidental. Intelligent vehicles and robots share significant overlaps in underlying technologies, including cameras, LiDAR, computing platforms, as well as perception, spatial modeling, localization, and planning algorithms, offering certain opportunities for technology reuse.
Automakers and autonomous driving suppliers have accumulated years of R&D experience in these fields and established relatively mature engineering development and supply chain systems. Robotics also provides new technical application scenarios for the automotive industry chain.
A mature intelligent driving system requires long-term investment in chip adaptation, sensor development, algorithm training, simulation testing, and vehicle validation. The technical assets formed by enterprises in these segments do not necessarily serve only passenger vehicles; they can also be extended to other intelligent devices.
Additionally, robotics offers autonomous driving companies a new market direction. Specifically, humanoid robots, whose physical structure is adapted to human-used spaces, tools, and facilities, may be easier to deploy than dedicated equipment in scenarios where environment modification is unsuitable.
However, this morphological advantage does not automatically translate into efficiency advantages. Suitability for a specific task still depends on manipulation capabilities, operational stability, and overall cost.
For automakers, existing manufacturing capabilities also hold value. Supply chain segments such as motors, electronic controls, batteries, sensors, controllers, and precision manufacturing intersect with the robotics industry.
The quality management, cost control, and mass production experience formed over decades in the automotive industry may help reduce robot manufacturing costs. However, components like robot joint modules, reducers, drivers, dexterous hands, and tactile sensors have their own technical requirements. The scaling capability of the automotive supply chain cannot replace the R&D and validation of these specialized technologies.
At the global launch of its Mojia robot on October 21, 2026, Chery disclosed that its AiMOGA (Mojia) robot business reuses approximately 70% of its R&D technology, with products and services entering more than 60 countries and regions. XPeng's layout goes further, as its robotics business involves not only product R&D but also the establishment of a relatively complete software and hardware technical system.
XPeng disclosed that the IRON robot has 76 degrees of freedom throughout its body and 21 degrees of freedom per hand, utilizing three self-developed Turing AI chips, achieving effective edge-side computing power of 2,250 TOPS. In August 2026, this business completed financing of over USD 900 million, with a post-money valuation exceeding USD 6.3 billion.
Autonomous driving supply chain companies are also attempting to extend into robotics. ZhiXing Technology's subsidiary, EmoStar Robot, began operations in March 2025, focusing on robot applications and launching robot controller products. Desay SV Automotive released the AI Cube, an AI computing terminal for the robotics field, in November 2025, adopting the company's technical architecture from the intelligent driving domain.
Although the entry points for these layouts differ, they all reflect that the autonomous driving industry chain is attempting to expand computing platforms, control technologies, and engineering experience into the robotics field.
However, while OEMs, autonomous driving algorithm companies, and Tier 1 autonomous driving suppliers are all accelerating their entry into the robotics field, their activities are not entirely identical.
OEMs often possess resources for whole-vehicle R&D, hardware manufacturing, and supply chains; algorithm companies' strengths lie more in perception, decision-making, data, and software platforms; Tier 1 autonomous driving suppliers may enter through products like computing platforms, controllers, and sensors. The reusable capabilities vary among different enterprises.
02 How Much of Autonomous Driving's Technical Accumulation Can Actually Be Reused?
From a technical architecture perspective, autonomous driving and embodied AI indeed share many similarities.
Autonomous driving systems need to acquire environmental information via sensors, identify roads, vehicles, pedestrians, and obstacles, predict the movement trends of other traffic participants, plan future trajectories, and finally execute decisions through steering, acceleration, and braking systems.
Robots also need to complete similar processes: acquiring environmental information, identifying objects and spatial relationships, understanding task requirements, planning actions, and executing movements via joints and actuators.
Both involve perception, decision-making, and execution, and both require handling continuous decision-making problems in uncertain environments. Among these, perception capability is the segment where reuse value is most easily visible. Autonomous driving companies have already accumulated extensive technologies in multi-sensor fusion, object detection, scene understanding, spatial modeling, and dynamic target tracking.
These capabilities can help robots understand their surroundings, identifying object positions, shapes, and motion states. Especially in relatively structured scenarios like factories and warehouses, visual perception, 3D spatial understanding, and target tracking experience accumulated in the automotive field provide a basis for transfer. However, robotic manipulation requires higher understanding of local object geometry, contactable areas, and pose changes. Identifying the position of a box is different from determining where to grasp it and how to grip it stably.
Traditional road object detection technology cannot directly solve the latter. There are also intersections in decision models. Behavior prediction and path planning in autonomous driving require the system to select appropriate actions based on environmental changes; robots need to determine the next action based on task goals, object states, and environmental constraints.
In recent years, the development of Vision-Language Models (VLM), Vision-Language-Action Models (VLA), and World Models has provided new technical paths for robots to understand complex tasks and generate action sequences. The autonomous driving field is also exploring stronger end-to-end decision models. While there are more intersections in model training and environment understanding between the two, specific tasks and control methods still show obvious differences.
Simulation training and data closed-loops also offer reuse potential. Autonomous driving companies have long relied on simulation environments for scenario generation, policy training, and extreme case testing. Robots also need to train repeatedly in virtual environments to reduce the time and cost of real-device testing.
Existing simulation infrastructure, model training platforms, and data management experience can serve as the engineering foundation for robot R&D. However, robot simulation must also handle physical processes such as contact, friction, collision, joint constraints, and object deformation. Autonomous driving simulation platforms can provide engineering experience, but new physical simulation models tailored to robot tasks are still required.
It should be noted that similarity in technical architecture does not mean identity in actual tasks. Autonomous driving systems primarily handle vehicle motion issues in road environments. Although vehicles constantly adjust speed and direction, their main control objective remains whole-vehicle motion, requiring comprehensive consideration of speed, acceleration, yaw motion, tire adhesion, and vehicle dynamics.
Robots are different. They not only need to move but also grasp, transport, rotate, push/pull, and assemble objects. Every operation may change the object's state and is affected by friction, contact position, object weight, and external constraints. This necessitates that robots possess stronger contact perception, force control, and fine manipulation capabilities.
For example, when facing a cardboard box, an autonomous driving system needs to judge its position, size, and motion state to decide whether to bypass it. For a robot, identifying the box is just the start of the task. It must also determine if the box is graspable, where to apply force, how much clamping force is needed, and how to prevent slippage during transport. This is no longer a simple problem of object recognition and path planning, but involves continuous interaction of vision, touch, force sensing, and motion control.
03 From Movement to Manipulation: Technical Difficulty Is Not a Simple Extension
When transferring autonomous driving technology to robots, the difference between motion control and physical interaction is most easily underestimated. Automotive control mainly revolves around whole-vehicle motion and vehicle dynamics. Systems control vehicles through steering, driving, and braking, needing to handle factors such as vehicle mass, wheelbase, steering characteristics, and road adhesion.
Although motion responses vary across different vehicle models, the primary control object remains the whole vehicle. Humanoid robots, however, need to simultaneously control multiple joints. Taking bipedal robots as an example, walking requires continuous adjustment of body center of gravity, support foot position, joint torque, and body posture.
If ground friction conditions, load, or external disturbances change, the original motion may need to be readjusted. If the robot also needs to perform grasping tasks simultaneously, control difficulty increases further. Arms, fingers, and the body need coordinated motion, and real-time perception of contact status is required when grasping objects.
Robots cannot decide finger movement solely based on camera-detected positions; they also need to use tactile and force feedback to judge whether the object is truly grasped and if the applied force is appropriate.
This indicates that trajectory planning experience from the autonomous driving field cannot directly replace the whole-body motion control and dexterous manipulation capabilities required by robots. Different robot tasks have varying requirements for manipulation capabilities. Tasks like transporting standardized boxes or inspections may only require relatively simple actuators; tasks like assembly, insertion, and flexible object handling impose higher demands on force control, tactile feedback, and fine manipulation.
Both types of systems must meet strict safety requirements, though the risk profiles differ. Autonomous driving mainly faces risks related to vehicle motion, road traffic, and collisions; humanoid robots must also address issues like pinching, collisions, falls, and unintended forces caused by close contact between joints, end-effectors, and humans or objects.
Especially for humanoid robots, actuators, joints, and dexterous hands may come into direct contact with humans. The system must not only know what task to complete but also ensure that action force, speed, and contact mode remain within safe ranges.
Therefore, while autonomous driving companies can reuse some perception, model training, and engineering validation capabilities, the robot body, actuators, dexterous hands, tactile sensors, whole-body control, and safety mechanisms still require establishing specialized technical systems.
04 The Real Bottlenecks Are Data and Scenarios
If technical architecture determines whether autonomous driving companies can enter the robotics field, then data and practical applications determine whether they can truly build competitive advantages. Autonomous driving has formed a relatively mature data closed-loop.
Vehicles collect information such as images, point clouds, and vehicle states during operation. Enterprises then extract valuable training samples through trigger mechanisms, data screening, and scenario mining, performing annotation, training, and simulation validation, and improving driving performance through model updates.
Robots also need similar data closed-loops, but data types and collection methods differ significantly. When completing manipulation tasks, robots need to record not just what they saw, but also what actions were taken, how much force was applied, what changes occurred in the object, and whether the action ultimately succeeded. For example, training data for grasping a cup cannot only include the cup's image and position but must also include finger contact positions, gripping force, object sliding situations, and final stability.
For tasks like screwing bolts, inserting wire harnesses, or folding clothes, contact feedback and operation results during the action process are equally important. For repetitive tasks, records must include whether the action succeeded, where failures occurred, and whether the robot could adjust the next operation based on feedback. This type of data is harder to collect and standardize than pure visual data.
Significant differences also exist in robot hardware. Arm lengths, joint layouts, dexterous hand structures, and degrees of freedom vary across manufacturers, meaning the same action might be executed completely differently on different robots. Even with large amounts of operational data collected, data adaptation issues between different bodies must be resolved.
A report on embodied AI published by MERICS in April 2026 pointed out that embodied AI training requires multimodal data support, while the cost of collecting operational data for robots in real physical environments is high; differences in hardware and software interfaces across different robot platforms also increase the difficulty of data sharing and model migration. This means that the massive amount of road data previously accumulated by autonomous driving companies cannot directly become robot training data.
What is truly valuable is migrating existing data engineering, simulation, and model training capabilities, and then establishing new data collection and training systems centered around robot tasks. Similar issues exist in the commercialization of embodied AI.
A report released by Counterpoint Research in August 2026 showed that global shipments of humanoid robots exceeded 22,000 units in the first half of 2026, a year-on-year increase of nearly 300%. However, looking at the application structure, humanoid robots used for entertainment performances, data production, and scientific research still account for over 60% of total shipments; smart manufacturing and warehousing logistics account for approximately 13% and 5%, respectively.
This indicates that rapid growth in humanoid robot shipments does not equate to large-scale productivity formation in factory and logistics scenarios.
A considerable portion of products is still used for demonstrations, research, and data collection. Shipment volume reflects product delivery scale, not necessarily stable, efficient autonomous operation in real production tasks. Therefore, for embodied AI enterprises, the robotics business cannot rely solely on technology demonstrations and product launches to establish long-term advantages.
Enterprises also need to find application scenarios with sustained procurement demand, proving that robots can stably complete tasks and meet customer requirements in terms of efficiency, cost, and reliability.
05 Final Thoughts
From XPeng increasing investment in robotics, to Chery attempting to reuse its automotive R&D system, to ZhiXing Technology, Desay SV Automotive, and other autonomous driving supply chain companies extending into robotics, the automotive industry chain is expanding its existing perception, algorithm, computing power, and manufacturing capabilities to new application scenarios. However, the advantages of autonomous driving companies mainly lie in environmental perception, spatial understanding, predictive decision-making, simulation training, data engineering, and scalable engineering management.
These capabilities can lower some barriers to robot R&D but cannot automatically fill gaps in motion control, tactile perception, dexterous manipulation, and physical interaction. Moreover, starting points differ for various enterprises.
OEMs possess supply chains, manufacturing systems, and quality management experience but still need to resolve specialized issues regarding robot bodies, actuators, and manipulation capabilities; algorithm-centric autonomous driving companies, despite accumulating expertise in perception, decision-making, and data training, need to further supplement hardware design, motion control, and mass manufacturing capabilities; Tier 1 autonomous driving suppliers can enter through computing platforms, controllers, and sensors but similarly need to adapt technical solutions based on robot tasks.
No matter which type of enterprise, the ultimate goal is to build a complete capability set ranging from robot body design and model training to actual task execution, rather than simply reusing a driving model.