2026 Changsha 1024 Programmers' Day: Embodied intelligence enters the market validation phase

On October 24–25, the 2026 Changsha 1024 Programmer Day and AI Builders Conference was successfully held at Changsha Zhigu. With the theme "Embodied Together, Building the Future with Intelligence," the conference points to a track shift that is already underway: over the past decade, AI's core application scenarios were concentrated in the "information domain," where we taught machines to read, write, compute, and converse; over the next decade, application scenarios will extend into the "physical domain," where machines must learn to reach out and manipulate, walk and move, turn screws, and care for humans. Embodied intelligence is no longer just a laboratory spectacle, but an industrial coordinate that technologists must now see clearly.

At this conference, more than a dozen speakers working at the industrial frontier shared their views on multiple dimensions including world models, real-machine data, dexterous hands, full-machine mass production, and even production-line deployment, channel delivery, and talent cultivation. When all their remarks are connected, the overall thread of embodied intelligence becomes clearer instead: it no longer depends on a single-point breakthrough in any one technology, but is a complete chain from "able to think" to "able to operate" and then to "able to be practical"—where the industry currently stands, how the brain iterates, how data accumulates, how the body is built, and how interaction with humans becomes reliable must ultimately return to the propositions of industrial deployment and "people" themselves.

01 "Entering the market validation period in 2026"

At the conference's Beijing·Changsha Embodied Intelligence Summit Forum, Analysys president Liu Yi offered a judgment: the global embodied intelligence industry will enter a market validation period in 2026 and will gradually enter real scenarios at scale. She used "four poles" to describe the global division of labor—the United States holds the "brain" of computing power and models, China holds the "body" of manufacturing and supply chains, Europe deepens the "rules" of standards and certification, and Japan and South Korea guard the "precision" of core components; the essence of the four-pole pattern is the misaligned distribution of "manufacturing capability, intelligence capability, rule-making capability, and precision capability." The differences in development paths between China and the United States are also clearly visible: the United States takes original breakthroughs in the "intelligence stack" as its core driving force, while China achieves inclusive development through large-scale deployment of the "manufacturing stack." She believes that companies with full-machine delivery, data closed loops, and deep scenario expertise will be the first to prove out business models, while companies relying only on single-point algorithms or demo narratives will face cash-flow tests. She recommends concentrating resources on batch validation and scenario data accumulation rather than chasing concept hype.

But as manufacturing capability becomes widespread with the diffusion of supply chains, industrial premiums are quietly concentrating toward the "brain" side. "Whether the brain is capable enough" is becoming the key yardstick determining whether an intelligent agent can truly handle complex tasks and occupy a core position in market competition.

02 "World models have become the second-stage thrust of the robot 'rocket'"

Wang Jianzong, chief technology officer of Huilun Technology and vice chairman of the China Open Source Software Development Alliance for Artificial Intelligence, likened world models to the second-stage thrust of the robot "rocket." His judgment follows a clear logic of capability upgrading: robot performance iteration is driven jointly by three stages, and the three core elements of world models, together with the four major characteristics of its self-developed world model, directly determine whether a robot can make effective predictions in scenarios it has never encountered. In concrete engineering deployment, Huilun Technology adopts a knowledge distillation route in which large models guide small models: using a model with a larger parameter count as the "teacher," it distills the implicit knowledge contained therein into a lightweight model, enabling the small model to achieve intelligence performance comparable to the teacher model at a scale far smaller than the teacher model. On this basis, it is further compressed and deployed to edge devices, achieving the effect of "operable offline, millisecond-level action command output, and continuous operation for 50 hours." Today, world models are gradually becoming a core candidate solution for robot operating systems.

03 "From 'understanding the world' to 'transforming the world'"

But world models must rely on real-machine data and simulation data of unified scale. The competition over world models is essentially a competition over the data foundation.

English text is being prepared.

04 "Models must learn to act, and should know 'what will happen after doing it'"

For models to truly master action capability, the most core investment is not computing power.

Mo Yilin, co-founder and chief scientist of Lingyu Intelligence and a tenured associate professor at Tsinghua University, believes that the core shortcoming of physical AI is not the model itself, but the absence of causal data on "what consequences an action will produce." In the field of embodied intelligence, the most valuable asset is data, and its cost logic is the exact opposite of the internet era—cost is not determined by annotation labor hours, but by hardware depreciation. He calculated that when the average unit price of the body is 120,000 yuan, the effective data collection cost is about 105 yuan per effective hour; when the body price rises to 600,000 yuan and the equipment intact rate drops to 85%, the collection cost soars to 640 yuan per hour. The cost difference mainly stems from fluctuations in body price and intact rate. Lingyu Intelligence's solution is to build a "one body, multiple brains" data infrastructure: turning commercial scenarios into continuously operating data factories, and developing long-tail problem correction, states before and after takeover, and system recovery processes into deliverable data products. Without feedback from the real world, models will remain confined to the simulation level.

05 "Let everyone have a personal robot"

No matter how elegant the data loop is, it ultimately has to be installed in a physical body that can walk and move.

Li Xiangdong, president of the Qiyuan Robot product line, titled this speech "Let everyone have a personal robot." His core logic is: embodied intelligent robots will become the next mass-scale intelligent terminal after mobile phones and automobiles, and landing in the consumer market depends on full-chain in-house R&D capabilities—the self-developed QDD quasi-direct-drive joint has a peak torque density of 85 N·m/kg, joint force control accuracy of 9.6% FS, reliability verification achieving 1 million wire harness bending cycles, and a mass production pace of one unit rolling off the line every 2.5 minutes. From industrial scenarios to personal consumer scenarios, many difficulties stand in between, such as cost, reliability, and scenario adaptation, but this also marks the industry's endpoint of development.

06 "Judge 'contacted' and 'can be stably picked up' separately"

The joints and drives of the whole machine body solve the question of "whether it can move"; but once it enters real scenarios, the tests of "whether it moves accurately and holds steadily" all rest on that pair of end effectors.

Su Yang, co-founder of Lingxin Qiaoshou, shared a pragmatic set of engineering judgment criteria: treat "has contacted the object" and "can stably grasp" as two independent dimensions for judgment. The dexterous hand that Lingxin Qiaoshou focuses on is the core limb closest to practical application and most technically challenging in the implementation of embodied intelligence—grasping, holding, twisting, and inserting, each action requires the coordinated interplay of force feedback, tactile perception, and robust control. The core principle he proposed is: judging whether a grasp is stable must rely on tactile feedback data for evaluation; data collection must both accumulate effective operational experience and fully retain error samples. In short, the core gap between robots and "truly skilled use of hand operations" lies in the lack of a tactile closed-loop mechanism and error-correction awareness, not in more demonstration effects.

07 "Bionic-skin robots within 1 meter"

Stable grasping means a single-step action has achieved reliability; but for embodied intelligence to integrate into human life scenarios, it needs to coexist with people "within 1 meter" and also complete long tasks of hundreds of steps coherently.

Liu Zikang, founder and CEO of Guangzhui Interactive, anchored the distance scale for human-robot collaboration at "within 1 meter." Using a simplified reliability calculation formula, he derived: under the model assumptions that each step is independent, the single-step success rate remains constant, and there is no retry recovery mechanism, the success probability of the entire task equals the single-step success rate raised to the power of the number of steps—when the single-step success rate is 99%, the success rate of completing a 100-step full-process operation is only about 36.6%; even if the single-step success rate rises to 99.9%, the full-process success rate is only about 90.5%. He said this is the industry's rational return to "long-task reliability and complete delivery capability." Within the 1-meter collaboration range, robots are not meant to replace humans, but to undertake collaboration, fill gaps, and do dirty and tiring work in scenarios humans cannot reach—but the premise for all this is that it must first cross the order-of-magnitude gap in reliability.

08 "Automation is the preset and solidification of production modes"

When robots can reliably complete long tasks by people's side, the intelligence of factories and production lines is no longer a pipe dream. According to data from the International Federation of Robotics (IFR) "World Robotics 2024," in 2023 China's manufacturing robot density reached 470 units per 10,000 workers, ranking third globally and surpassing Germany and Japan; in 2024, China's newly added industrial robots accounted for 54% of the global total. This industrial foundation is precisely the confidence for embodied intelligence to land in manufacturing.

Deng Wenping, chairman and CEO of Viewtrix Robot, proposed that automation is essentially the preset and solidification of production modes and cannot autonomously adapt to batch adjustments and disordered incoming materials—this is precisely the natural gap between automation and intelligence. The solution he offered is "Kunwu Platform—reconstructing a new paradigm of equipment manufacturing with physical AI": Kunwu integrates digital twins, industrial large models, and low-code generation to build an intelligent foundation for production lines where "simulation is development, and design is delivery"; in virtual space it completes 3D production line construction, virtual commissioning, process verification, and control system generation, then deploys them to the real workshop with one click, realizing a perception-decision-execution closed loop. The foundation is supported by the factory simulation platform KUNWU and the machining world model KWorld. Relying on this capability, the market share of the construction machinery blanking and sorting production line has reached 80%. For embodied intelligence to land in manufacturing, the ultimate winner is not a single robot, but a complete physical AI brain that can be reused across scenarios.

Technology can run on the production line, but someone still has to deliver it to the customer—this is the other half of industrial deployment.

09 "Bridging the 'last mile' of embodied intelligence applications"

Xiao Xiangjiang, founder and chairman of Hunan Chaoneng Robot, president of the Hunan Artificial Intelligence Association, focused on channel building and system integration. The "Chaoneng Robot World Industrial Internet Platform" he led the creation of has opened the world's first robot 9S store offline, expanding to 20 cities in just half a year, with its core positioning precisely being to "bridge the 'last mile' of embodied intelligence applications." In his view, embodied intelligence currently faces non-standard long-tail pain points: enterprises lack independent R&D capability, manufacturers struggle to sell products, users find it hard to use equipment well, and local governments find it difficult to attract supporting enterprises—the platform does not take over every link, but rather takes on non-standard demands and then transforms them into replicable standardized modules. A set of intelligent brains capable of cross-industry migration relies precisely on the engineering organization capability behind it to complete re-parameterization for each customer site.

And when hardware, software, production lines, and channels are all stacked together, the hardest constraint surfaces: people. No matter how smart a robot is, people still have to define scenarios, break down tasks, and conduct acceptance checks.

10 "But first you have to teach the teachers"

Wang Ting, director of the Education Business Department of Hunan Chaoneng Robot Technology Co., Ltd. and deputy general manager of Hunan Chaoneng Intelligent Technology, shared her insights starting from education deployment. Her view is very straightforward: an individual mastering AI application capabilities can replace an entire team, but the first prerequisite is to teach teachers how to use AI first. She provided a set of shocking data: nationwide AI compulsory courses are expected to require 1.2 million full-time teachers, while there are currently fewer than 20,000 full-time teachers, a gap of 98%. The deployment path she leads is divided into two parts: at the K12 stage, starting from textbooks to create a full-chain closed loop; at the university level, laying out training laboratory construction and talent incubation. For embodied intelligence to achieve large-scale development, it does not rely on a single enterprise's demonstration prototype, but on a whole cohort of professionals capable of transforming industry scenarios into specific tasks and then breaking those tasks down into acceptable actions.

In addition, at the roundtable dialogue of the Changsha·Beijing Embodied Intelligence Summit Forum, Xiao Bujuan, director of government-enterprise cooperation at Songyan Power; Jiang Weilai, professor at the College of Artificial Intelligence and Robotics of Hunan University and director of the Hunan University CETC Industrial Embodied Intelligence Joint Laboratory; and Zhang Rui, deputy general manager of Zhiyu Rongxiao, supplemented aspects of this value chain from the perspectives of humanoid robot bodies, industrial embodied intelligence platforms, and university-industry-research collaboration.

Closing remarks

From the technical exchange at the foot of Yuelu Mountain in 2020 to the 2026 Changsha Zhigu venue with more than 2,200 attendees, over seven years, what has changed is the development tools, the number of sessions, and the scale of the events. What has never changed is that there are always a group of practitioners willing to gather together to discuss the true value yardstick of technology.

What this conference presented is not only technology and products, but also a calibration of industry coordinates. From data foundations and world models to end effectors, from complete machine delivery and channel building to talent cultivation, embodied intelligence has moved from a conceptual puzzle to a stage where progress is measured by scales—each step corresponds to clear engineering metrics and commercial validation. As "degree of intelligence" replaces "hardware parameters" as the focus of competition, the industry is also shifting from chasing a single technology to systematically considering complete machine delivery, data closed loops, and scenario deepening capabilities. This is the clearest signal of this programmers' festival: the scale of embodied intelligence is set by deployment, not by stories.