56. A Conversation with Five Embodied-Intelligence Companies: The Robot’s “Brain” and “Body”—Which Is Waiting for Which?

Summary

• The 2025 guests described embodied AI with “proliferation,” “emergence,” “siphoning,” “changing by the day,” and “mass production”; the host added “the year of breakout growth.” 蔡颖鹏 said multiple robot OEMs shipped more than 1,000 units this year, while industry-wide shipments likely exceeded 10,000 units, with component shipments also above 10,000 units; he noted that this may be his own estimate. 姚卯青 said embodied AI is siphoning talent and capital from autonomous driving, language models and vision models, making it the “grand synthesis” of these technology paths. 唐文斌 said the sector is accelerating and faith in technology is growing firmer, but it is also “quite frothy.”
• Whether hardware is the bottleneck depends on the scenario, perspective and time horizon. 刘东 said the brain may be the constraint in material-handling scenarios, while hardware may constrain dexterous or fine manipulation. 唐文斌 distinguished research from deployment, arguing that hardware is not the bottleneck from a technology-development perspective, but that application reliability and stability require hardware development time that cannot be compressed. 姚卯青 sees algorithms as the bigger constraint for now: tasks teleoperation can complete remain beyond algorithmic coverage in both cycle time and success rate, although over the long term hardware remains far from human levels of dexterity, energy density and reliability. 蔡颖鹏 also leans toward software as the larger bottleneck.
• 唐剑 divides the bottlenecks into “linear” and “nonlinear” categories. Hardware issues such as joint heat, torque density, dexterous-hand degrees of freedom and size, on-device compute and batteries could make substantial progress each year with sustained investment. But when large-model generalization will reach its “ChatGPT moment” is impossible to forecast: it could happen next year, or in 3, 5, or even 10 years.
• “Models unlock scenarios; scenarios define hardware” is an important frame, but the answer differs by layer. 唐文斌 said the scenario determines the hardware form factor: physical limits such as power density mean light- and heavy-payload applications cannot both be optimally served by the same configuration. 刘东 distinguished between the upper-layer multimodal brain and the lower-layer VLA: the former can be deployed across robot bodies through standardized sensor configurations, while the latter must be tightly coupled to the body. 姚卯青 stressed that “who defines whom” and “who constrains whom” are separate questions; hardware also co-adapts with customer scenarios. 蔡颖鹏 likewise said scenarios, models and hardware should continuously feed back into one another.
• The industry should pursue closed-loop value, outcome orientation and common standards. 唐文斌 wants products to create customer value in real-world settings that can be calculated, measured and ultimately deployed at scale, while questioning “framework orders” that lack a real value loop. The RoboChallenge evaluation platform was initiated by 原力灵机 and Hugging Face, with 智源研究院, 智源机器人, 自变量, 星海图, Tsinghua University and Xi’an Jiaotong University among its participants. 唐剑 called for common standards covering joint interfaces, data-collection formats and evaluation, while leaving model and joint-module architectures free to proliferate.
• The guests broadly agree that humanoids are not a detour, but they are not the whole answer. 唐文斌 said “humanoid is a means, not an end”: humanoids are best on emotional appeal, but not necessarily on functional value. 刘东 said specialized tools do not always need to be operated by humanoid robots; giving the tools their own brains can be more effective. 唐剑 said general-purpose humanoids may, like general-purpose PCs and smartphones, fully or partially replace some dedicated equipment, while heavy-payload industrial robots will remain for the long term. 姚卯青’s team has already pursued wheeled dual-arm, bipedal and quadrupedal forms. 蔡颖鹏 sees humanoids as one of the most important embodiments of core technologies and the broader industrial chain.
• The 2026 agenda includes on-device compute, scaled deployment, data growth and a possible emergence moment. 刘东 said large models must run on the robot itself and expressed hope for high-performance domestic edge chips; when asked about related initiatives, he confirmed that “it’s already underway.” 唐文斌 is betting on Scale across data and applications. 姚卯青 said some overseas companies have already accumulated data on the order of several hundred thousand hours and cited π0.6’s use of reinforcement learning and other autonomous methods to collect real-world interaction data, arguing that next year may bring an emergence moment not seen before. 唐剑 sees scaled humanoid deployment in vertical sectors as relatively certain, but the embodied-AI “ChatGPT moment” remains highly uncertain.

Deep dive

1. 2025’s Keywords: Proliferation, Emergence, Siphoning, Change by the Day and Mass Production

• 蔡颖鹏, speaking from the supply-chain side, picked “mass production”: multiple robot OEMs shipped more than 1,000 units this year, while total industry shipments likely exceeded 10,000 units, with components also above 10,000 units; he added, “This data may be my own estimate.” In his view, the industry’s cumulative volume over the past several decades may not have been as large as this year’s alone, laying the groundwork for data collection and exploration of real-world use cases.
• 姚卯青 chose “siphoning”: embodied AI is drawing talent and capital away from autonomous driving, language models, vision models and other fields, making it the “grand synthesis” of these technology paths after years of development.
• 刘东’s keyword was “proliferation,” although he believes the industry’s technology paths, solutions and robot bodies have multiplied without the overall deployment path converging. 唐文斌 chose “emergence,” saying the world’s pace has accelerated and conviction in technology has strengthened, while the practical pursuit of value remains “quite frothy.” 唐剑 described the sector as “changing by the day”: new robots, motion-control behaviors and embodied algorithms are advancing almost daily. Host 卫诗婕 added that when she visited Silicon Valley in October, local investors and hardware founders envied China’s supply chain.

2. Whether Hardware Is the Bottleneck Depends on the Scenario, Perspective and Time Horizon

• 刘东 answered by scenario. Material handling may be done with a forklift, leaving the bottleneck in the brain’s ability to recognize the environment and task. Dexterous or fine manipulation is different: the algorithms may already have a first-generation model, while the hardware still cannot execute the task.
• 唐文斌 separates research from application. From a technology-development perspective, hardware is not the bottleneck: today’s grippers are already quite good, but they remain far from the intelligence humans display when using them. In applications, however, hardware remains a bottleneck because a product must operate continuously, reliably and stably. That requires hardware-development time that cannot be compressed; it cannot simply be hacked together at software speed and put directly into service.
• 姚卯青 sees algorithms as the bigger bottleneck at this stage. The share of tasks that can be completed algorithmically remains low relative to what teleoperation can achieve, both in cycle time and success rate; algorithms still cannot cover everything the hardware can do. Over the long term, however, hardware remains far from human levels of dexterity, energy density and reliability.
• 蔡颖鹏 also leans toward software as the larger constraint. Many demos can be made to work through debugging, teleoperation and similar methods, but models and vertical-industry data are still missing when systems enter actual sectors. Some scenarios remain beyond both the software and hardware.

3. Linear vs. Nonlinear Bottlenecks: 唐剑’s Framework and 蔡颖鹏’s Supply-Chain Checklist

• 唐剑 calls hardware a “linear bottleneck”: joints still run hot and torque density remains low. A dexterous hand gets larger as degrees of freedom rise; with fewer degrees of freedom, it may be possible to keep the hand small using 6 degrees of freedom, but size and performance are hard to optimize together. On compute, some demo robots trail an Nvidia RTX 4090 behind them; after Orin, the industry is still waiting for Thor or more powerful domestic chips. Battery energy density remains low: solid-state batteries are better, but their cycle life is still nowhere near that of liquid-electrolyte batteries. Sustained capital and headcount investment could produce substantial annual gains in these hardware areas.
• The true “nonlinear bottleneck,” he says, is software—specifically, when large-model generalization will reach its “ChatGPT moment.” It could happen next year, or in 3, 5 or even 10 years.
• 蔡颖鹏’s component list includes motors, reducers, lead screws, bearings, sensors, compute, rotary joints, linear joints, end effectors, tactile sensors and 6-axis force sensors. These products have made substantial progress in technology, maturity, cost and consistency in recent years, but if the underlying materials and physical principles do not change, they will still hit ceilings quickly. Further breakthroughs will require advances in foundational technologies such as materials and industrial machine tools.
• His concrete example: when a factory moves loads weighing tens of kilograms, the passive load-bearing structure of a dexterous hand may be able to handle the load, while the arm’s active load capacity cannot; the problem is joint torque. As the number of degrees of freedom rises, motors must be packed more tightly, and the constrained space reduces the load capacity of each joint. Higher force output also brings faster heat buildup.

4. Who Defines Whom: Models Unlock Scenarios; Scenarios Define Hardware

• 唐文斌 rejects the idea that the model defines the hardware or the hardware defines the model; the scenario defines the hardware. Physical limits such as power density mean light- and heavy-payload applications cannot both be optimally served by the same form factor. Otherwise, the machine will likely be overkill and too expensive. Models unlock scenarios, and scenarios then define hardware in return（模型解锁场景，场景定义硬件）.
• 刘东 gives a layered answer. At the top, the multimodal brain defines the hardware to a greater extent: lidar, vision sensors and audio input can be standardized into a sensor configuration and deployed across different robot bodies. The VLA at the lower, cerebellar layer should be tightly coupled to the body. Cross-body VLA models are possible, but performance will be better when the model is tuned to a specific class of robot body.
• 姚卯青 first separates “who defines whom” from “who constrains whom.” At this stage, models plainly cannot define hardware; once hardware is taken into a real-world scenario, the customer’s requirements define and refine it. During deployment, even a cooperative customer’s production line may adapt to constraints such as the robot’s reachable workspace and cycle time, so scenario and hardware remain in constant interaction.
• The industry’s rule of thumb that hardware sets the floor and software the ceiling still holds（硬件决定下限，软件决定上限）, 唐剑 says, but software and hardware should not be split into a sequential handoff from the robot body to motion control and then to algorithms. A robot is not necessarily better when it is lighter; the weight ratio between the upper and lower body also matters. That is the kind of requirement a software team can feed back to the hardware team, creating a closed loop between them.
• 蔡颖鹏 likewise sees scenarios, models and hardware as mutually reinforcing. Scenarios and models are driven more by the user’s perspective, while hardware seeks the best engineering solution; approaching the problem from only the scenario or model side may overlook physical limits.

5. What the Industry Should Demand: Closed-Loop Value, Real-Robot Benchmarks and Interface Standards

• 唐文斌 calls for a “genuine value loop”: once a product enters a real-world setting, it must create customer value that can be calculated and measured, and ultimately be deployed at scale by customers. It is “very dirty, very hard and very exhausting” work. What he rejects are “framework orders” without a clear scenario or value validation, asking whether they actually close the loop and whether the scenario value is real.
• The host said that a few days earlier she had hosted the open-source release of Pelican, the embodied-brain model from 北京人形, calling it the largest open-source embodied-brain model by size. RoboChallenge was then formally released. The platform was initially launched by 原力灵机 and Hugging Face, with 智源研究院, 智源机器人, 自变量, 星海图, Tsinghua University and Xi’an Jiaotong University among the organizations jointly building it. 唐文斌 wants the industry to define common standards and evaluation methods, because only by knowing how to measure performance can the industry know how to improve it.
• 唐剑 draws the line between what should be standardized and what should remain open. Embodied models and joint-module architectures can proliferate, including planetary, harmonic, cycloidal and linear pushrod designs. Joint interfaces, data-collection formats and evaluation standards, however, should be unified; otherwise the industry will waste resources. He also believes open source will have a profound impact on the sector.
• 姚卯青 warned that the sector’s rise could bring a flood of companies, homogeneous competition and unhealthy behavior from marketing through product quality. The industry should avoid these problems as much as possible and focus on fair evaluation, customer value, commercial closure and results that ultimately show up in company revenue.

6. Humanoids Are Not a Detour: A Means, Not an End

• 唐文斌’s formulation is that “humanoid is a means, not an end”（人形是手段，人形不是目的）. Humanoids are the best form from an emotional-appeal perspective, but not necessarily from a functional one; both forms of value are driving robotics forward.
• 刘东 says humanoids are certainly not a detour, but embodied AI will not be limited to humanoids. Humans have learned to use tools, and having a humanoid operate a tool is sometimes less effective than giving the tool its own brain so it can complete the task autonomously.
• 唐剑 believes that humanoid robots with general-purpose embodied capabilities “will be a historical inevitability.” They may resemble the general-purpose PCs of the 1980s and the general-purpose smartphones of the first decade of this century, fully or partially replacing certain dedicated devices such as MP3 players or PDAs. Heavy-payload industrial robots, however, have already optimized single-point applications and efficiency and will remain for the long term.
• 姚卯青 does not expect the industry to go all-in on one form factor and wait until it proves unworkable before reconsidering. Taking 追觅机器人 as an example, he said the team has pursued wheeled dual-arm, full-size and half-size bipedal, and quadrupedal forms covering companionship, service and material-handling scenarios.
• 蔡颖鹏 expects embodied AI to support multiple robot forms, but sees the humanoid as one of the most important—possibly the single most important—because it brings together core technologies and the broader industrial chain. It therefore will not be a detour.

7. 2026 Outlook: On-Device Compute, Scale and the Emergence Moment

• 刘东 sees on-device compute as 2026’s biggest challenge. The industry cannot rely on cloud support to bring embodied AI into even some deployment scenarios; large models must run inside the robot itself. Products such as Nvidia Thor have already appeared, but domestic compute remains relatively far behind. When the host asked about related initiatives, he confirmed that “it’s already underway.”
• 唐文斌 boiled his expectations down to one word: “Scale”—scaling data, scaling the actual deployment of use cases and scaling applications across more companies.
• 姚卯青 expects an emergence moment in embodied AI once data volumes reach a sufficient threshold. He recently saw some overseas companies reach the several-hundred-thousand-hour range in data. He cited π0.6, which uses reinforcement learning and other autonomous methods to continuously collect real-world interaction data, and said next year may bring the first signs of emergence moments not seen before.
• 唐剑 sees scaled humanoid deployment in several vertical sectors next year as relatively certain, while continuing to look for the embodied AI “ChatGPT moment”—an outcome he calls “highly uncertain.” 蔡颖鹏 likewise sees vertical industries and vertical applications as the key commercialization areas to watch next year.