2026-08-19 · Xiangming Technology
In August 2026, Unitree Technology closed up 460.34% on its first day of listing on the A-share market, with its closing market value reaching 340 billion yuan in one fell swoop. The pricing given by the capital market with real money has far exceeded the scale that a robot dog or humanoid robot hardware manufacturer should have. What truly deserves the attention of software practitioners is not how fast or how high this machine can run or jump, but another competition hidden behind the halo of hardware—the battle over operating systems and software ecosystems for embodied intelligence. This is repeating the historical inflection point of smartphones shifting from "Nokia's hardware is king" to "iOS/Android ecosystem is king."
Humanoid robots have gone through a critical "hardware-first" stage. Over the past five years, the industry's narrative focus has been almost entirely concentrated on hardware links such as mechanical structures, reducers, joint motors, and dexterous hands. Reducer precision, motor torque density, and whole-machine degrees of freedom constituted the core indicators by which capital and media evaluated a robotics company.
But after manufacturers such as Unitree, Figure, and Tesla Optimus pushed hardware parameters to a certain threshold, an awkward fact surfaced: hardware is no longer the bottleneck—software is. Between a robot that can walk stably, grasp, and go up and down stairs and one that can truly "work" lies an entire software stack—from low-level real-time control and motion planning, to mid-level perception fusion and task decision-making, to upper-level scenario-oriented application orchestration.
The smartphone industry has provided a clear frame of reference. When the iPhone was released in 2007, touchscreens and multi-touch were not unique to Apple; what truly set it apart was iOS and its app ecosystem. Today's humanoid robots are standing in a similar position: whoever first builds a reusable, scalable, low-threshold robot software platform will control the distribution rights of the next decade.
Deconstructed from a technical perspective, the software stack of embodied intelligence can be roughly divided into three layers, each undergoing independent architectural evolution.
The bottom layer is the real-time control and motion planning layer.This layer has strict requirements for latency and determinism, and usually runs on a real-time operating system (RTOS) or a Linux kernel patched for real-time performance. The coordinated control of dozens of degrees of freedom across the whole machine, force feedback loops, and gait and grasping planning all need to be completed within millisecond or even sub-millisecond cycles. Traditional robot manufacturers used closed proprietary controllers, but a clear trend over the past two years has been migration toward "general-purpose Linux kernel + real-time extensions + standardized middleware," and message middleware such as ROS 2 and DDS is becoming the de facto standard.
The middle layer is the perception-cognition-decision layer.This layer is the main battlefield where large language models (LLMs) and vision models (VLMs) exert their strength. Multimodal large models have given robots the ability to "understand scenes and interpret tasks in natural language" for the first time, and the "brain" of embodied intelligence is shifting from rule-based state machines to unified vision-language-action (VLA) models. But the engineering reality is that there is a natural rate mismatch between the latency of model inference (usually hundreds of milliseconds) and the real-time requirements of low-level control (millisecond level). How to build a software architecture for hierarchical decision-making and asynchronous scheduling is one of the most difficult technical challenges at present.
The top layer is the developer-facing application and toolchain layer.This is the core of ecosystem competition and the layer most like an "operating system." Whoever can provide a complete SDK, simulation training environment, hardware abstraction interfaces, and development and debugging tools can lower the development threshold and attract more third-party developers to build scenario applications on this platform—just as Android attracted app developers. At present, this layer is still in a fragmented state: manufacturers each go their own way, lacking unified application standards and distribution mechanisms.
If we look at the competitive landscape, there are at least three technical routes competing in parallel.
The first is the route of "body manufacturers building their own ecosystems."Represented by Unitree and Tesla, they possess the strongest hardware definition power and massive real-world operating data, and tend to make the software stack a moat for their own hardware. The risk lies in closedness—historically, if a closed software ecosystem does not have overwhelming market share, it is often eventually dismantled by open ecosystems.
The second is the route of "open-source neutral platforms."Represented by the ROS 2 community and various open-source VLA frameworks, it advocates making the robot software foundation a public good and having all hardware manufacturers connect to it. This path best serves developers' interests, but commercially it lacks a clear monetization loop, and its progress is constrained by the efficiency of community collaboration.
The third is the route of "cloud vendors/large model vendors extending downward."Platform companies such as Baidu, Huawei, and Tencent are trying to package cloud-based large model capabilities, edge inference frameworks, and industry solutions and extend them downward into the robotics field. Their advantage lies in computing power and models, while their weakness lies in lacking a deep understanding of a physical body.
It is worth noting that these three routes are not a zero-sum game and are more likely to move toward "layered decoupling"—standardization of low-level control middleware, cloud-based supply of mid-level model capabilities, and upper-level applications led by scenario parties. This is precisely the space where software outsourcing development and industry solution companies can enter: most traditional enterprises do not need to develop their own robot operating systems; what they need is to package mature robot capabilities into applications in specific scenarios, such as inspections in smart communities and sorting in logistics warehouses. For such customers, choosing a software development team familiar withIoTand industry application integration is often more pragmatic than building a technology stack in-house.
One clear judgment is that the focus of competition in embodied intelligence is irreversibly shifting from "body hardware" to "software ecosystems." Hardware parameters will gradually converge, just as today's flagship phone screens and chips can hardly constitute differentiation; ultimately what determines share will be software experience and the developer ecosystem.
For software practitioners and enterprise customers, this means two specific lines of action. First, do not overbet on any single piece of robot hardware; instead, place capability building at the software layer—understanding general technologies such as ROS 2, multimodal model inference, and edge deployment is what allows technical assets to remain reusable as bodies iterate rapidly. Second, prioritize entering through "light scenarios" for validation, using the smallest hardware investment to run through a business loop, and then decide whether to scale. FromXiangming Technologybased on its experience serving local enterprises in Shenzhen, customers willing to first validate in small-scale scenarios and then scale up have a noticeably higher success rate in implementation. Technology itself is not the goal; using technology to create real business value is the judgment least likely to become outdated amid this clamor.
Shenzhen Xiangming Technology Co., Ltd. | Create value with technology | xiangmingit.com
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
