Gasgoo Munich-"The moat for embodied intelligence isn't just a single model or one robot—it's technology, scenarios, and a continuous closed loop of data and interaction." Those words from Wu Xiang, general manager of Pudu Robotics' embodied intelligent product line, captured the industry's core mix of anxiety and anticipation at the 2026 World Artificial Intelligence Conference.
AI is stepping out of the digital realm, and robots are no longer content with just performing backflips on exhibition stages. A fundamental question is emerging: Do we really need to train a dedicated "brain" for every unique robot shape? Or can we build a universal intelligence foundation that allows different "bodies" to share the same "mind"?
"One brain, multiple forms" has emerged at this industry crossroads as one of the most closely watched directions in embodied intelligence. The concept aims to let robots of different shapes share a single "brain" through a unified underlying model and software architecture.
Image source: Pudu Robotics
But is "one brain, multiple forms" a smooth path to the future, or merely an over-idealized technological utopia?
From "One Machine, One Brain" to "One Brain, Multiple Forms"
"One machine, one brain" used to be the unspoken norm across the robotics industry.
A floor-cleaning robot ran on algorithms dedicated to path optimization; an industrial mechanical arm relied on a control system built solely for repetitive precision. Experience rarely flowed between them, data couldn't be shared, and algorithms were incompatible. Every time a new robot form was introduced, the software and algorithms had to be built almost entirely from scratch.
The industry calls this "chimney-style development"—silos rising up with little connection between them. Transferring algorithms across different forms has long been a persistent efficiency pain point.
But change is underway.
Recently, global consultancy Frost & Sullivan released its "2025 Global Embodied Intelligence and Commercial Service Robot Independent Market Research Report." It notes that the robotics industry is evolving from "one machine, one brain" toward "one brain, multiple forms."
As application scenarios expand, the report says, traditional models can no longer support cross-scenario replication and large-scale deployment. "One brain, multiple forms" connects diverse robot bodies through a unified intelligence base, enabling different robots to share capabilities in perception, task planning, skill invocation, and data loops. The future of the industry lies in shifting from standalone product delivery to platform-based, ecosystem-driven, and infrastructure-like development.
"One brain, multiple forms" isn't a concept created out of thin air; it's a consensus formed as the industry seeks scalable implementation. As Wu Xiang put it: "General embodied intelligence isn't about using one configuration to solve all tasks. It's about building a universal intelligence foundation, then deploying different types of robots based on the task at hand."
The core logic is this: by building a unified underlying model and software architecture, robots of different forms—wheeled delivery bots, bipedal humanoids, or fixed-base mechanical arms—can share the same "brain." This brain handles perception, understanding, planning, and decision-making, while the different "bodies" handle execution. Intelligence is reused across forms, eliminating the need to train from scratch every time the hardware changes.
It aims to resolve a triple dilemma that has long plagued the industry: models lack generalization and fail easily in new scenarios; high-quality physical-world data is scarce, and collecting real interaction data is far harder and costlier than scraping internet text; and every new scenario or hardware change results in low reuse rates for algorithms and data, forcing massive amounts of work to be redone.
Li Yan, deputy director of the Industrial Economy Research Department at the Development Research Center of the State Council, noted in an interview that large-scale commercialization of embodied intelligence requires coordination on three fronts: models must have cross-scenario generalization; supply chains must be mature and reliable; and a data flywheel effect must be established in real-world scenarios. Generalization is prioritized first, as it solves "the universality of the same product across different scenarios, aiming to achieve economies of scope."
The industry is moving in this direction. According to the "China Embodied Intelligence Industry Development Report (2026)," China's embodied intelligence market is projected to hit 1.09 trillion yuan in 2026, with a compound annual growth rate exceeding 22%. Driving this growth is demand across multiple industries and scenarios for general robotic capabilities—where cross-form reusability is a critical link.
A joint initiative by the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission on humanoid robots and embodied intelligence field training is also pushing robots to accumulate multi-form real-world data in industrial, service, and special operations scenarios. That data is precisely the necessary foundation for training a "universal brain."
From the silos of "one machine, one brain" to the universal foundation of "one brain, multiple forms," the direction is becoming clear.
Who's Betting on It?
With the direction set, the path forward becomes the critical question.
Pudu Robotics is a prime example. Starting as a commercial service robot maker, it now operates four product lines—delivery, cleaning, and industrial—covering 16 industries. Its business spans 85 countries and regions with cumulative shipments exceeding 130,000 units. According to Frost & Sullivan, it ranks first globally in both revenue and shipments for commercial service robots, holding roughly a 44% share of China's service robot export market.
Leveraging this first-mover advantage, Pudu began pondering a deeper question: could the "mobile intelligence" accumulated in delivery and cleaning scenarios be transformed into a capability reusable across different forms? The answer is its three-layer Physical Agent architecture.
Image source: Pudu Robotics
The Body Layer—specialized robots handle high-frequency, standardized tasks; humanoid-like robots enter complex industrial and commercial environments; and humanoid robots explore more general human-robot collaboration scenarios. All three forms share the same intelligence foundation.
The System Layer—PuduAgent OS, a general embodied agent platform responsible for task understanding, planning, decision-making, experience accumulation, and safe operation. Wu Xiang likens it to "the Android and iOS ecosystem of the physical world."
The Skill Layer—PuduFM, a proprietary embodied foundation model with a built-in physical intuition module. Trained on Pudu's own dataset, which it claims is the world's largest for robot physical interaction, it emphasizes genuine understanding of physical space. Robots must not only "see" objects but also "understand" their gravity, friction, and inertia.
The synergy of these three layers allows robots to achieve a complete closed loop of "perception—understanding—planning—execution—feedback" in real-world scenarios.
EBKernel positions itself as an embodied intelligence brain provider. Its Cognitive World Model can currently adapt to various robot forms, including bipedal, quadrupedal, and mechanical arms.
Addressing the challenge of multi-body adaptation, Liu Jinyu explained that the issue can be broken down into two core dimensions: drive structure and movement form. Differences in underlying drive components—electric, cable, hydraulic—directly affect adaptation logic. Meanwhile, bipedal, quadrupedal, and wheeled platforms all move differently.
With numerous robot manufacturers and no unified industry protocol, interoperability remains difficult. In Liu's view, a single embodied intelligence brain cannot yet achieve full-category adaptation across the entire industry at this stage.
Given this landscape, EBKernel adopted a focused adaptation strategy. It anchored core partners across various forms, covering quadruped, wheeled, electric, and cable-driven configurations. During model pre-training, the team collects data from different robot configurations in the same scenario. This diverse coverage hones the model's cross-body generalization, rather than blindly expanding boundaries by partnering with every possible hardware type.
Mech-Mind's core thesis is that different tasks and scenarios will spawn different robot forms—robot shapes won't converge. However, core intelligence capabilities like perception, understanding, planning, and execution can be continuously accumulated, copied, and migrated across different robot types.
Consequently, Mech-Mind has built a general "eye-brain-hand" technical system around "one brain, multiple forms." Centered on embodied large models and a biomimetic layered robot brain, it combines industrial-grade 3D vision, motion planning, robot control, and end-effector capabilities to form a complete closed loop.
Image source: Mech-Mind
Relying on this unified system, Mech-Mind can configure these intelligent capabilities on demand for different robot forms. Currently, its "eye-brain-hand" setup supports humanoid, industrial, mobile, and collaborative robots.
Rather than simply migrating general large language models or vision-language models into robotics, Mech-Mind started from real robot tasks. It built a universal intelligence foundation for the physical world, focusing on environmental perception, task understanding, autonomous planning, and motion execution.
There are many similar cases. JUEJANG Technology demonstrated another variation of "one brain, multiple bodies." A company born from collaborative robots, it showcased its platform at a recent seminar, using its proprietary "Kongyi" large model to drive mechanical arms, humanoids, and other forms.
During WAIC, ACE Robotics unveiled its "Xiaotu" quadruped all-domain operation solution. Composed of an embodied super-brain module A1 and an intelligent control platform, it flexibly adapts to various quadruped robot bodies and has already been deployed in scenarios like security patrols at Caohejing Park and urban management along the Shanghai Bund.
These cases show that "one brain, multiple forms" has become industry consensus, though paths diverge. Some start with the data flywheel, others with brain-like intelligence, and still others extend from strengths in perception and manipulation. The starting points differ, but they point to the same goal: letting one brain master multiple bodies.
Consensus and Challenges
The direction is set, but how far is reality? Between consensus and actual implementation, several hurdles remain.
Liu Jinyu, co-founder and COO of EBKernel, was blunt in a recent speech: "Unlimited cross-body adaptation is indeed costly. From joints down to whole-machine interfaces, the cost of heterogeneity is much higher than imagined. Limited cross-body adaptation is feasible, but unlimited adaptation isn't commercially realistic."
In his view, adapting one model to nearly 200 OEMs and roughly 400 products on the market is neither realistic nor necessary.
Data is another hurdle. Large language models can grow quickly by scraping internet text, but a robot's "brain" needs high-quality physical interaction data. Every run in a real environment, every physical interaction, must be collected via hardware deployment. More troublesome is that "mobility" data isn't the same as "manipulation" data. The entire industry faces the same problem: getting robots to move is easy, but getting them to work like humans is still a long way off.
The technology itself is difficult. Cross-body generalization remains a common challenge; from single-point manipulation with dexterous hands to whole-body motion control, a host of technical problems still need solving.
"We're working toward this goal, and we think there's a chance," Liu said. "Of course, it's localized, out-of-the-box generalization right now, not fully plug-and-play. It's hard to reach everyone's expectations, but I think it's a direction worth striving for. At least we can see it's feasible—it's not completely impossible. It will take time, and we're trying to push it forward with our technical capabilities."
This is the prevailing view among industry practitioners: challenges exist, but the direction is clear. The 2026 WAIC sent a clear signal: the race in embodied intelligence has officially shifted from a "physical fitness contest"—counting how many push-ups a robot can do—to a "brain power contest" focused on continuous learning and self-evolution.
True "one brain, multiple forms" won't arrive suddenly on a single day. It will emerge in layers and evolve continuously across different scenarios like factories, supermarkets, and homes. Ultimately, the players who can simultaneously build strong models, manage data well, manufacture capable hardware, and keep operations running in the real world will be the ones to define the final rules of this game.
