Written by | Wu Xianzhi Edited by | Wang Pan
This month, the embodied AI industry has seen continuous major developments. In just half a month, Noetix Robotics has officially released its embodied AI sub-brand Scalabot and launched two model products.
On September 8, it released its first world model, the HERON-World Model. A week later, Scalabot unveiled the second model under the HERON technical architecture system—the HERON-Context Reinforcement Action Model (CRA). Its core capabilities include memory, reinforcement learning, and cross-embodiment generalization.
When asked why the company is entering the "brain" field at this time, Jiang Zheyuan, founder and chairman of Noetix Robotics, stated that embodied AI suffers from the wooden barrel effect, and the current bottleneck has shifted from the physical body to the brain. With a hint of end-game thinking, he believes that low-cost, scaled physical bodies are a prerequisite for data-driven scaling. The closed loop of body + brain is the mainstream path in the era of data flywheels, and the industry's endgame will likely see a "winner-take-all" scenario.
Although the technological roadmap has not fully converged, there is a preliminary consensus on the paradigm of VLA + world models.
In addition, relying on its extreme craftsmanship and physical body capabilities, Noetix Robotics' physical body business is about to achieve self-sufficiency, giving it the surplus capacity to build the brain.
Besides the discussion on entering the brain field, there have been many hot topics in the embodied AI industry recently. For instance, the boss of a newly listed embodied AI company pointed out on his WeChat Moments that some companies are using data collection orders to inflate their revenue.
When asked the same question, Jiang Zheyuan stated that Noetix Robotics has not accepted a single data collection project so far. Meanwhile,
Recently, media outlets including Photon Planet conducted an in-depth communication with Jiang Zheyuan, bringing a series of questions. We have edited the content of this communication without altering the original meaning of the dialogue. Below are the main highlights of this interview:
Now Is the Right Time to Build the Brain
Q: When did Noetix Robotics decide to enter the brain field? Why exactly this year?
A: The team was formed in July last year and only gained combat effectiveness this year. The reason for entering the field is very simple: the right person came along, and everything fell into place naturally. In addition, we have accumulated enough this year to showcase it externally.
Q: Who is this person? How large is the current team?
A: I can't say. Let me check the team size on Feishu, (looking at the phone) dozens of people.
Q: How did you persuade him to join? Did you offer an irresistible quote?
A: Basically, everyone who has talked to me has joined. The team is very young, and the work we are doing can attract talent. The company will absolutely not offer an irresistible price, you can rest assured about that.
Q: Does entering the brain field mean you have seen the convergence of the technological roadmap?
A: I didn't say the technological roadmap has fully converged. It's just that the route of world models + VLA is relatively clear. It is still too early to say it has completely converged.
Q: There are already dozens of companies building brains in the market. Isn't it too late to enter now?
A: Physical body companies and brain companies arrive at the same destination via different routes; both are building a general-purpose embodied AI robot, just with different paths. Embodied AI has a "wooden barrel effect," where the amount of water it can hold depends on the length of the shortest plank.
In 2022, apart from Boston Dynamics, no humanoid robot body in the world could walk stably. The physical body and motion control were the shortest planks. Today, the physical body is ready. At this point, the next bottleneck to address is the brain. Previously, the technological roadmap had not yet converged. We waited until the roadmap relatively converged and a certain consensus was reached before taking action.
The brain cannot exist independently of the physical body, nor can the data itself. Data drives everything. To achieve Scaling, data needs to be At scale (massive scale) and Diverse (extremely diverse). Therefore, physical body companies with scaled, low-cost manufacturing capabilities will have an incomparable advantage in Scaling.
Q: When you say the physical body is ready, does it mean a passing grade or a high score? Or has the performance already supported us in building an excellent brain?
A: I want a passing grade. A physical body must meet several characteristics: First, it should not be a consumable that fails in a short time; it must last at least one or two years. Second, it should have the ability to perform operations, including sufficient degrees of freedom, large enough torque, and small enough joint backlash. Third, the underlying motion control must support complex operations.
If NVIDIA's SONIC or BFM (Behavioral Foundation Model) can support operational tasks, it reaches the passing grade. Which is better for the remaining 80 or 90 points? There are different evaluation metrics for this. There are many answers, such as more sensors, full-body skin tactile sensing, or stronger reliability. There is no final conclusion yet.
Q: How do you view the leading advantage in data accumulation of those brain companies that have been in the field for two or three years?
A: Brain companies have thicker cash reserves, but judging from the current results, brain companies are explorers of the entire industry. They have paid a lot of sunk costs, and there are not many capabilities that can be precipitated to generate compound interest.
Q: What pitfalls do you think physical body manufacturers have avoided?
A: We do not comment on everyone else's choices; let's talk about our choices.
First, in terms of data, we choose a multi-layered complementary data pyramid: Ego, UMI, simulation, and multi-body real machines.
We have made a judgment: all major data sources inevitably require trade-offs across multiple dimensions. Data must be highly heterogeneous and diverse, and there is no complete homogeneous substitutability between data. This is our core judgment on the direction of scaling.
Second, at the model level, we choose to explore world models and VLA in parallel, and combine the two using reinforcement learning.
There is still debate over VLA and world models. The HERON-World Model is responsible for "rehearsing the future," and the HERON-CRA is responsible for "executing actions"—allowing each route to be explored deeply enough rather than rushing to merge them. But parallel does not mean disconnected: the reinforcement learning engine combines the two into the same framework. The world model is integrated into the RL framework of CRA, providing the value model with deductions of "what will happen if this action continues," helping the policy judge whether the action is good or bad; the successful and failed experiences generated by real machine execution flow back to drive the next round of iteration. Thus, the robot obtains a complete closed loop of "understanding, deduction, action, and error correction."
Third, we choose a route to "make the data flywheel truly spin."
The victory or defeat in this industry is not in demo videos, but in the real physical world. We take the initiative in every link of the flywheel: self-developed physical bodies enable low-cost, scaled data collection; the data pipeline forms a full-link closed loop from disk storage, cloud uploading, annotation to training and evaluation; problems with the model are exposed on the real machine and can be quickly fed back to data and training.
Where Is the Data Inflection Point?
Q: In the early stage, whether it was world models or VLA, there was sufficient discussion. Unitree also sparked some controversy early on because it didn't build a brain. How does Noetix Robotics persuade investors?
A: Persuasion is far less important than choice. Some investors simply do not look at physical body companies and only focus on the brain, while other investors only look at the physical body. A company must have at least one business that can support cash flow first before talking about the stars and the sea. For companies that purely build large models, business revenue can be quantified by metrics like ARR. But embodied AI companies that only build brains do not have this, and it is hard to see returns in the short term. Therefore, I think we must build the physical body first.
Q: What scale of data volume do you think is required for embodied AI to reach the capability inflection point?
A: At least on the scale of millions of hours, but I think millions of hours are highly likely not enough.
Q: Is the main source of the millions of hours self-collected or massively outsourced?
A: We will have a ratio. Ego data is very general and mainly relies on external sources, while physical body data is entirely our own.
Q: The promotional information mentions that the robot goes from knowing how to do it to learning and doing it better and better. Is this repetitive training for a single task?
A: Some prerequisites need to be defined. Generalization refers to the ability to complete tasks with a very high success rate when handling unfamiliar scenarios and tasks. According to this definition, there are two lacking points. First is the lack of data scale and data diversity. Second is the lack of On-policy Data.
Q: From your perspective, can the current data production and collection fully meet the demand?
A: The current industry still needs to categorize different types of data. Our own idea is to put more energy into physical body data. The quality of Ego data still has room for further improvement because quality depends on the quality of data annotation and pipeline efficiency.
Q: Model companies can buy some cheap physical bodies as materials and then collect data. Where is Noetix Robotics's advantage?
A: Our cost is low, but we don't necessarily sell it cheaply; there will be a financial gap here. In addition, the key issue is not just the physical body, but a complex set of solutions, involving systematic capabilities such as how to collect data, whether there is dedicated equipment, how to transmit data back, and how to annotate it.
I think that apart from the physical body cost, other parts are replicable, it's just that we haven't seen this happen yet. We warmly welcome foundation model companies to purchase physical bodies on a large scale. To quote a great man, "Success need not be mine, but I must contribute to it."
Q: Everyone is talking about the next decade of embodied AI. What do you think companies will compete on after entering the second half?
A: I think it's about competing in data infrastructure, who can obtain more high-quality, diversified, and scaled data the fastest.
Q: Noetix Robotics' world model looks like a general platform for data input and command output. Does this mean you will open it to other manufacturers in the future?
A: The HERON-World Model is not a platform for inputting data and outputting commands. It is an action-conditioned world model. Given the current visual frame and the action the robot is about to execute, it deduces how the world will change next. The actions that are actually outputted are from the CRA.
Regarding openness, let's separate two things. We do not plan to turn the brain into a platform service. Our energy is focused on making the model stronger and our own physical body products better. This is the core of Noetix Robotics' closed loop of body and brain.
However, we agree with technology inclusion, which is also the brand positioning of Scalabot. Therefore, we will consider sharing some technological achievements with the industry in an open-source manner, allowing research institutions, developers, and industry partners to use them for research and verification, and continue to innovate on top of them. Embodied AI is still in its early stages. Working together to build a solid foundation is more valuable than each company reinventing the wheel behind closed doors.
The world model can serve as a policy evaluator and data generator, reducing the cost of real machine production. Moreover, we have trained it on multiple physical bodies and multiple types of data, giving it a certain degree of universality.
Q: Some pursue data scale, while others pursue data quality. Which do you think is more important?
A: At this stage, quality is more important, and quality includes diversity and annotation accuracy.
On Cost Competition: Before Scaled Deployment, Cost Is Always a Key Variable
Q: Is Noetix Robotics' cost control based on scale?
A: It relies on craftsmanship. There is a very reasonable saying that R&D accounts for 80%, while the supply chain and scaling only account for 20%. To reduce costs, craftsmanship must account for over 80%. Dropping from CNY 25,000 to CNY 23,000 can rely on scale, but dropping from CNY 100,000 to CNY 20,000 must rely on craftsmanship. We have used a lot of craftsmanship, and the team is helping the entire industry step into pitfalls.
Let me give a few points. First, the entire physical body of the whole machine is made of all-plastic, and no one in the entire industry has replicated it now. All-plastic means that the physical body structure cost will be significantly reduced compared to metal materials.
Currently, there is no solution in the world that is as reliable and low-cost as Noetix Robotics's joints. I take responsibility for my conclusion.
Q: Noetix Robotics has a very strong supply chain, but building a brain is a completely different logic. How do you align internally?
A: Building a brain and cost are complementary. Let me give an example. There are two types of companies in autonomous driving. One is Waymo, which does not look at the cost of a single machine, uses the best components, and stacks up autonomous driving. The other is Tesla, which uses the most ordinary hardware and solutions to minimize costs in exchange for data scale, and data scale can exchange for very strong intelligence.
To this day, Tesla will soon be able to break into the Robotaxi market, while it becomes very difficult for Waymo to replicate it. Therefore, lowering costs is a prerequisite for data-driven approaches.
Q: Do dexterous hands need to be designed by ourselves?
A: They are all externally procured. High degrees of freedom come with high costs, while low degrees of freedom are insufficient, which is a pain point. However, we currently have no plans to make them ourselves.
Q: Small humanoids are selling well today. Can this last until tomorrow?
A: Looking back at the previous wave of similar products, consumer-grade quadrupeds, growth has already peaked today. This shows that user pain points need to be solved, not just Early Adopters. There is a "Crossing the Chasm" theory. The first batch of users must be Early Adopters, and there is an uncrossable chasm between them and the mainstream mass market, which is satisfying rigid demands.
When the growth of consumer-grade quadrupeds peaks and the entire market turns downward, looking at our consumer-grade small humanoid market, we think we need to consider more: who will be the next batch of buyers? Penetration will always peak. Only by crossing the chasm and entering the mainstream market can we truly scale up.
Q: What will the next-generation product answer?
A: Next year.
Q: Compared to the products of current competitors, the price of making bionic products will be much higher.
A: We are particularly strong in building physical bodies and very good at reducing costs. At that time, the cost will definitely be far below the industry level, but higher than the existing product line.
Q: How do you view the inevitable trend of price reduction for robot physical bodies?
A: You can look at Unitree's financial report. The main force driving revenue is the humanoid G1. I don't think it's a price reduction, but rather opening up some new price bands. Unitree G1 hasn't reduced its price much and is still a hard currency, so I don't see a price reduction. What I see are more robot products in various different price bands.
After we released Bumi, some companies may indeed follow up and release products in similar price bands. I think they are still far from delivery, and this is not a price war.
Short-term Verbal Battles and Long-term Industry Competition
Q: Recently, the competition among several brain companies in the industry has been quite fierce. For example, a listed company attacked an unlisted one, and then three other companies complained about being reported. What do you think?
A: We really want to organize a gathering, invite everyone to have a good cup of tea together, and stop fighting each other. This is meaningless. In the early stage of industry development, we must never fight each other. We still focus on our own development and do not comment on our peers.
Q: The recent verbal battle has focused on data collection orders. Will Noetix Robotics have them?
A: Our company has not had a single data collection order to date. We almost took one before, but didn't take it in the end. Once we saw the wind blowing the wrong way, we gave it up. We are not short of revenue either.
Q: Do you think data collection orders are definitely inferior orders? What is the main source of revenue for Noetix Robotics?
A: It depends. Some data collection involves buying physical bodies for data collection, which are real orders. But if they buy your physical body and also ask you to buy the data back, that's not a real order. The bulk of our revenue this year comes from projects, such as B2B industry-education integration and educational informatization projects.
This benefits from past accumulation, and signs of scaling have emerged this year. In addition, there are commercial service scenarios, which are quite diversified. Because we found that in many solutions, partners will do secondary development for more scenarios based on our robots.
Q: At the humanoid robot sports meeting, we saw the gap between manufacturers widening. Will many companies disappear because of this next?
A: In F1 racing, have you ever seen the Wuling team? No. Have you ever seen the Great Wall team? None of us have. NIO was one of the founding teams of Formula E in its early years, and now it no longer participates.
Racing events must be in a specific historical stage to boost industry development. When the technological increment and commercial value created by the event are lower than the enterprise's expectations and cannot make the ROI work, it will be like a rocket launch, shedding the previous stage's booster.
Therefore, I don't think participating or not is related to a company falling behind. Lotus Cars have been very successful in various races, but commercially, it is not a very successful enterprise. Later, after going through several twists and turns, it was successively acquired by Proton and Geely. Who still remembers that it used to be a noble British aristocrat?
Q: Is your money enough? If you are short of money, what is the main area?
A: If it's just about surviving, there is definitely no problem. If you put CNY 10 billion in my account today, I can deploy 10,000 robots to collect data. I might be able to achieve L1+L2 very quickly. The most money-burning part is the brain, and the cost is several times higher than the physical body and motion control.
Q: If Bumi's profit is relatively low, why do you still make Bumi a very important product line?
A: Good question, here are a few points. First, brand mindshare needs to be accumulated over time. Let me share some practical experience. When we launched Bumi on e-commerce platforms, the store traffic was particularly low. Over time, it was gradually nurtured.
In the consumer market, brand mindshare must be nurtured over time. We need to continuously plant seeds, continuously make our voices heard, continuously deliver, and continuously generate UGC content from users, forming the cognition that Noetix Robotics must be chosen for humanoid robots. Therefore, even if the gross profit is low, or even the entire business may temporarily lose money, we must do it.
Second, Bumi in the consumer market will be of certain help to our brain R&D. We are currently working on self-developed models, and internally we have a middle-platform department called the Interaction Center to do model fine-tuning.
Q: Can you estimate the trend of the industry? In the long run, will it be a winner-take-all for the leaders, or will the leadership change every year or two like in the new energy vehicle industry?
A: Winner take all. It is more like the US new energy market. Because of the data flywheel effect, after the physical body goes on sale, the returned data is both On-policy and massive, highly diverse data. I think the home end will definitely be winner-take-all.
Q: What determines the leader? Technology, scaling, or organizational capability?
A: I think it's the route. In the US, actually only Tesla bet on pure vision end-to-end, letting users help collect data. Actually, many car companies in China are already doing this, but ultimately who can become the real leader depends on the sales scale of the cars and the data closed-loop capability. In the robot industry, it depends on who sells to homes first and sells explosively first; that company will be the outstanding one.
WeChat ID | TMTweb Official Account | Photon Planet