Author | Wang Han, Editor | Mo Ying
In 2026, capital is pouring investments into AI entering the physical world. Data from IT Juzi shows that as of May 11, 218 investments have been disclosed in the domestic embodied intelligence sector this year, with a total amount exceeding CNY 57.7 billion. In less than half a year, this has already surpassed the total of 357 deals for the entire year of 2025. Morgan Stanley has raised its shipment forecast for humanoid robots in China for the second time this year, increasing it from 14,000 units at the beginning of the year all the way to 50,000 units. The protagonists in this race are not just robots.
Looking up, CCID Consulting predicts that the market size of China's low-altitude economy will exceed CNY 1 trillion in 2026; looking down, the China Commercial Industry Research Institute estimates that the market size of China's autonomous driving will reach CNY 529.3 billion in 2026.
From NVIDIA founder Jensen Huang calling Physical AI "the next wave of AI," to Figure AI completing 1,250 hours of real production operations at a BMW factory, AI's march into the physical world seems unstoppable.
But what about reality? Morgan Stanley's breakdown shows that among the humanoid robots shipped in China in 2025, 42% went to R&D and education, 19% were used for interactive displays, 19% for data collection, and only 4% were actually working in industrial logistics scenarios. Money has been invested and goods have been shipped, so why can't physical intelligence be deployed at scale? On September 18, during Huawei Connect 2026, Huawei, jointly with the China Academy of Information and Communications Technology (CAICT), the Shenyang Institute of Automation of the Chinese Academy of Sciences (CAS), the Global Computing Alliance, and EY (China) Enterprise Consulting Co., Ltd., released the "Research Report on Cloud-Edge-Device Collaborative Architecture for Physical Intelligence." The core judgment is: Physical intelligence must move from "single-point bottlenecks" to "system evolution," and the path to system evolution is cloud-edge-device collaboration.
01.Why Does Physical Intelligence Need Cloud-Edge-Device Collaboration?
Before discussing the dilemmas of physical intelligence, we must first clarify: what is physical intelligence?
Huawei defines it in the report as the capability system of AI to complete the closed loop of perception, understanding, reasoning, planning, action, and feedback in the real physical world, with carriers including robots, autonomous driving, smart spaces, and industrial automation equipment. Drone route planning, real-time decision-making in autonomous driving, optimized design of industrial control systems, energy scheduling in buildings, etc.—as long as AI needs to perceive the physical world and take action on it, it belongs to physical intelligence. Although the forms of manifestation differ, they all face the same set of underlying architectural issues: the dilemma between "real-time performance" and "complexity."
Luo Yunfeng, Director of the Computing Domain in Huawei's ICT New Opportunity Incubation Department, states bluntly that "AI entering the device" is the biggest bottleneck for the implementation of physical intelligence. On the device side, current mainstream robot bodies are still subject to multiple constraints in computing power, power consumption, cost, and space.
On the cloud side, the inference latency of large models on the cloud cannot meet the millisecond-level requirements of industrial control. Meanwhile, the multi-dimensional perception of robot and drone bodies means massive uplink bandwidth, whereas traditional networks are primarily downlink-oriented. In weak or disconnected network scenarios, pure cloud solutions also face the risk of paralysis. More fundamentally, regarding the tasks themselves, the operating frequencies of the six layers of tasks—perception, decision-making, execution, collaboration, learning, and operations—vary significantly, making it impossible for a single node to form a closed loop. This dilemma reveals the gap between device performance and capability.
Zhai Guanglei, President of Huawei's ICT New Opportunity Incubation Department, gave an example: "Usain Bolt runs 100 meters in 9.58 seconds, and robots can do it in 9.36 seconds; humans can high jump about 2.4 meters, and robots can already jump over 3 meters this year. But robots even struggle to tighten screws on a production line, because precision manufacturing requires refinement down to 0.1 to 0.2 millimeters, requiring a combination of vision, precise touch, force control, and dexterous hands." Running fast and jumping high are performance; tightening screws steadily is capability.
Performance can be stacked up at single points, but capability must rely on a system. Since a single point cannot simultaneously satisfy real-time performance and complexity, the only solution is to let the device, edge, and cloud each perform their respective duties: the device side focuses on microsecond/millisecond-level real-time perception and high-frequency motion control, solving "speed" and "safety"; the edge side, as the regional hub, undertakes data aggregation, local model fine-tuning, multi-robot collaboration, and weak-network autonomy, solving "stability" and "connectivity"; the cloud side, as the global intelligent brain, is responsible for large model development, global planning, and operations, solving "power" and "intelligence".
Reducing reliance on single-point computing power through layered collaboration, while balancing deterministic latency, complex reasoning, multi-robot collaboration, and global optimization, is the systematic path for the large-scale implementation of physical intelligence.
02.Breaking "Three Walls," Connecting "One Network," and Turning "One Flywheel"
How exactly should cloud-edge-device collaboration be implemented? Huawei's "Research Report on Cloud-Edge-Device Collaborative Architecture for Physical Intelligence" proposes the core proposition of "Three-Domain Collaboration and One-Network Integration."
Three-domain collaboration corresponds to the three main flow lines of tasks, data, and computing power. Application collaboration allows a task to be reasonably split, scheduled, and collaboratively executed across the cloud, edge, and device; data collaboration completes the alignment of multi-modal data within the domain—spatiotemporal calibration of vision, force, and tactile sensors is indispensable, while also achieving cross-domain perception sharing and integration of training and inference data; computing power collaboration uniformly schedules computing power resources of different locations and paradigms, allowing developers to use computing power "seamlessly."
The premise for all this collaboration is that the three layers must be connectable and smoothly communicative, which is the task of "One-Network Integration." It covers the entire domain of physical intelligence with intra-device networks, inter-device networks, and device-cloud networks: the intra-device network ensures millisecond-level transmission from intelligence to electromechanical systems, the inter-device network supports point-to-point collaboration among multiple robots, and the device-cloud network is responsible for the two-way communication of data and instructions.
Technically, relying on the integration of TSN, 5G-A, NearLink, and industrial Ethernet buses, it builds full-domain connectivity covering "intra-device, inter-device, and device-cloud": NearLink, with microsecond-level latency and ultra-high reliability, is responsible for high-speed wireless interconnection between sensors and controllers inside the robot body; the industrial Ethernet bus (such as EtherCAT), with sub-microsecond-level jitter, is responsible for deterministic wired synchronization of precision equipment on the production line. 5G-A, with large uplink and millisecond-level latency, feeds back massive perception data to the edge or cloud, and issues global scheduling instructions to the site.
As a unified wired backbone, TSN provides unified deterministic scheduling and time synchronization for 5G-A backhaul traffic, industrial Ethernet bus traffic, and NearLink aggregated traffic, ensuring that all service flows work collaboratively under the same time reference without interfering with each other. The network is thus upgraded from "connecting devices" to "connecting tasks and intelligent capabilities," becoming the foundation of the entire architecture. As intelligence moves into the physical world, a massive amount of high-quality data is essential. Yao Maoqing, President of the Embodied Business Department of Agibot, mentioned in his speech at WAIC 2026 that the embodied intelligence industry generally faces "three walls": the data wall, the representation wall, and the closed-loop wall.
The solution provided by Huawei's report is the data flywheel of "real-machine data feedback—cloud training and evaluation—edge-device model deployment." Real-machine operation data flows back to the cloud for training and evaluation, and the optimized models are then deployed to the edge and devices, forming a spirally ascending capability loop. This simultaneously requires the collaboration of three types of data sources: real data, simulated data, and generated data—real data provides physical authenticity, simulated data provides scale, and generated data provides variants.
To make this architecture implementable, the report also clarifies four major design principles: layered decoupling and collaborative autonomy; real-time priority and intelligent collaboration; elastic scheduling and optimal resources; open compatibility and continuous evolution. When agents begin to take action in the physical world, the security mechanisms themselves must also collaborate across the cloud, edge, and device.
03.Cloud-Edge-Device Requires Gradual Implementation; Industrial Collaboration is the Ultimate Answer
The architectural solution is clear, but implementation still needs to proceed step by step. In the report, Huawei divides the cloud-edge-device implementation process into three stages:
- In the short term, the edge is deployed on demand, primarily focusing on device plus cloud, first running through scenarios such as retail and inspection;
- In the medium term (2028–2030), it moves towards regional autonomy, adding "one brain, multiple machines" on the edge side, and enhancing "one machine, multiple brains" on the device side;
- In the long term, it moves towards dynamic collaboration, forming an integrated distributed intelligent network.
Regarding implementation scenarios, Huawei judges that physical intelligence should first be implemented in structured professional scenarios such as manufacturing, logistics, energy, and commercial services, and then expand to open and complex scenarios such as home services and urban logistics. Wang Jianwei, Vice President of Huawei's Manufacturing and Large Enterprise Group, judges: "At this stage, embodied intelligence is very much like the early stage of autonomous driving in 2017 and 2018. This industrial cycle will still take 3 to 5 years before it can be implemented at scale." The "Three-Domain Collaboration and One-Network Integration" model has already produced a benchmark through pioneering exploration. Huawei and the National-Local Co-built Humanoid Robot Innovation Center officially released the first embodied intelligence training ground benchmark—the Qilin Training Ground at this year's WAIC, which is also the country's first heterogeneous humanoid robot embodied intelligence training ground.
The core of this benchmark is to turn "teaching robots to work" into an industrialized assembly line. The cloud-based Ascend ultra-large-scale intelligent computing cluster is responsible for pre-training, edge nodes support real-time inference, and low-power chips on the device side are responsible for the body's motion control. More crucially, the formation of the closed loop: after the model is deployed to the real machine on the device side, the cloud will continuously monitor the operating status, collect data and issues, enabling the real-machine data from the device side to flow back and feed the next round of training. The data flywheel of "real-machine data feedback—cloud training and evaluation—edge-device model deployment" mentioned in the report has completed its first large-scale practice here. Currently, robots trained from this system have entered scenarios such as automobile manufacturing, smart retail, urban sanitation, and special operations, verifying applications in the three major fields of industry, people's livelihood, and special operations.
04.Conclusion: System-Wide Collaboration and Continuous Evolution to Truly Bring AI into the Physical World
Physical intelligence is a complex systems engineering project spanning chips, foundation models, communication networks, real-world data, and industry implementation scenarios. To bring AI from the virtual digital space into the real physical environment and complete the closed loop of perception, decision-making, and execution, relying solely on the optimization of single-point scenarios and the iteration of single technologies can no longer bridge the complete technical chain, making it difficult to achieve large-scale implementation.
Huawei's "Research Report on Cloud-Edge-Device Collaborative Architecture for Physical Intelligence" points out a path worth exploring for the industry: "Three-Domain Collaboration and One-Network Integration." Currently, key technologies such as cross-layer collaboration and heterogeneous scheduling are still in the exploration stage. The competitiveness of physical intelligence will ultimately depend on the comprehensive capabilities of system-wide collaboration and continuous evolution.