VLA end-to-end models can understand instructions but are unable to answer "why do this": YaoShi Technology does not manufacture complete machines, but develops a brain-inspired core that supports cross-ontology reuse, is interpretable, and is capable of learning during operation.
Let intelligence grow like life — instead of manufacturing complete machines, we build a brain that can be reused across different robot bodies and keeps learning during operation.
Over the past two years, the humanoid robot track has never lacked public attention.
According to IDC data, global shipments of humanoid robots in the first half of 2026 reached nearly 25,000 units, a year-on-year increase of 432.1%, of which about 19,000 units were from China. Goldman Sachs predicts that global shipments will rise to 502,000 units by 2032; the Chinese Institute of Electronics forecasts that the market size of humanoid robots in China is expected to reach 870 billion yuan by 2030.
Capital is also pouring into this sector at an accelerated pace. In August 2026, Unitree Technology was listed on the Sci-Tech Innovation Board, with its opening market value once exceeding 440 billion yuan; according to statistics from Zhidongxi, as of June this year, there are at least 25 domestic embodied intelligence unicorns valued at over 10 billion yuan.
However, beneath the boom, the focus of competition is shifting — from "whether the robot can move" to "whether it can work stably, durably and interpretably in real scenarios".
This exactly points out a shortboard of the current mainstream technical route. Most embodied intelligence companies bet on the VLA (Vision-Language-Action) end-to-end model. It acts as a powerful "generalization engine" that enables robots to understand instructions and generate actions, but its decision-making process is a black box: it is hard to explain why the robot takes certain actions; the training relies on a large amount of teleoperation data with high cost; after the model training is completed, its capabilities tend to be fixed, and it has limited recovery ability when encountering failure states outside the training set.
The "body" part is developing rapidly, but the supply of "brains" lags behind relatively. Most leading companies deliver products in the form of complete machines or full-stack solutions, where the brain capabilities are deeply bound to their own robot bodies. For many complete machine manufacturers and integrators, "having a body but no matching brain" is a realistic dilemma. There are still very few independent brain suppliers in the market.
Shanghai Yaoshi Technology Co., Ltd. (hereinafter referred to as "Yaoshi Technology") has chosen a different path: instead of making complete machines, it focuses on developing a brain-inspired robot brain that does not rely on specific hardware. The project was conceived and launched in 2024, and two years of R&D investment have been made so far, with the system keeping running and iterating continuously. The company is currently seeking 5 million RMB in angel round financing.
Beyond the black box, an alternative brain development route
Yaoshi Technology defines its product as a "brain-inspired bionic robot brain", which is delivered in pure software form and can adapt to robot bodies of different manufacturers and different forms.
"Brain-inspired" does not mean "shape-simulated". The company does not focus on making robots grow a human face, but on the internal organization mode of intelligence: how living beings perceive the world, form memories and make judgments. Corresponding to the system, it is a layered brain region architecture built with reference to the anatomical structure of the human brain, and different brain regions are responsible for perception, memory, emotion, motion coordination, decision-making and other functions respectively.
The key difference from the VLA end-to-end model is that this architecture splits the cognitive process into traceable links: perception, abstraction, cognition, knowledge, memory, strategy, emotion and consciousness. All links have clear boundaries, and when the robot makes a judgment, the whole link can be traced back segment by segment.
The direct result brought by this is that the decision-making link is traceable: when the robot takes an action, you can trace back along the link to see what it has perceived, what concepts it has formed, which memory it has called, and what kind of strategy trade-off it has made. For high-risk scenarios such as medical care, elderly care and industry, this traceability is the basis of interpretability — "why the action is taken" is as important as "whether the action is correct".
In terms of memory and learning, Yaoshi Technology does not adopt the idea of "one-time training, long-term use", but integrates the learning process into the operation process. The system will continuously correct judgments according to the action results, and integrate new experiences into existing memories instead of restarting from scratch every time. Multiple learning modes are unified under the same learning architecture, sharing the memory and attention mechanism.
The company has also introduced a system-level decision-making mechanism called "energy balance", which is used to coordinate the conflicts between safety and autonomy, short-term benefits and long-term impacts. According to Yaoshi Technology, this is not a philosophical packaging, but an executable balance framework: the system continuously observes the overall operation state, and gives correction directions when imbalance occurs.
Another key capability is "cross-body adaptation". The brain can recognize and gradually adapt to the connected robot body, without needing to redevelop a new brain for each individual body. In other words, the brain is not bound to a specific robot body; even if the body is replaced, parts are modified or even damaged, the brain still has the opportunity to re-adapt. This is exactly the technical premise for the establishment of an "independent brain supplier".
Let the brain grow during operation
Whether the technical route is feasible ultimately depends on whether the system can actually run properly.
According to Yaoshi Technology, the company has completed multiple rounds of engineering verification and learning capability acceptance, and finished stability rectification; the new round of learning capability deepening has entered the operation period acceptance stage.
At present, the system has achieved 7×24 hours continuous operation, with stable capabilities including concept formation, relational memory (able to distinguish between acquaintances and strangers), circadian rhythm, prediction based on historical experience and adaptive attention. The company has also opened up two voice naming learning routes of "naming objects by sight, recognizing objects by hearing": the robot does not look up answers in a syllable table, but gradually maps sounds to objects through repeated coordination of seeing, listening and speaking.
These capabilities sound abstract, but they determine whether the robot can "get more familiar with the real environment the more it is used". The capabilities of traditional end-to-end models are basically finalized when leaving the factory; what Yaoshi Technology wants to build is a brain that accumulates experience and corrects judgments in the process of interacting with the environment.
On the other side of the progress, there is still a missing part that has not been completed yet. Up to now, Yaoshi Technology's brain has not completed joint debugging on the real robot body, and the existing verification is mainly completed in software environment and experimental environment. After financing, the company plans to cooperate with 1-2 complete machine manufacturers to complete the demonstration of the voice naming learning closed loop on the real robot body, and then promote the complete closed loop of perception-cognition-decision-action.
Intellectual property is a very important asset of this company at present. Yaoshi Technology says that it has deployed more than 100 patents, covering types of invention patents and utility model patents; a number of new invention points have completed patentability assessment and are entering the application preparation stage.
Instead of making complete machines, we build "to-be-integrated" brains
Yaoshi Technology's commercialization path is very clear: it will not compete with complete machine manufacturers, but authorize its brain products to complete machine manufacturers and integrators.
The revenue mainly comes from three parts: brain solution authorization, vertical scenario solution and customized development: authorize the brain SDK/API to complete machine manufacturers and charge annual authorization fees; provide "brain + integration + service" for scenarios such as scientific research and education, health care and companionship, and commercial exhibition; carry out secondary development for customers with special needs. In the long run, the company also hopes to cover the scientific research and education market through developer ecology and SDK to form standard influence. According to the company's current plan, the authorization fee will fluctuate with the scale of complete machine manufacturers and the complexity of scenarios, and the specific price is still in communication with the first batch of partners.
The target customers are also clear: one category is complete machine manufacturers and integrators that have humanoid or mobile robot bodies but lack advanced cognitive capabilities; the other category is industry customers with clear requirements for "interpretability, auditability and sustainable learning", such as elderly care institutions, special education schools, medical rehabilitation institutions and flexible manufacturing enterprises.
The implementation rhythm given by Yaoshi Technology is: first verify the developer ecology in the scientific research and education market, and at the same time create real robot body benchmarks with 1-2 complete machine manufacturers; then complete POC in 2-3 vertical scenarios; finally enter the large-scale authorization stage. The company judges that the real revenue growth will not come until the real robot body verification is completed.
The advantage of this path is "lightweight": it does not manufacture robot bodies, does not bear the supply chain pressure, and does not compete with customers for the complete machine business. The difficulty is also straightforward: the value of the brain must be reflected through the robot body, and the joint debugging of real robot bodies relies on partners; under the background of the rapid maturity of the VLA route, the brain-inspired architecture needs to produce comparable scenario results as soon as possible.
This company is also a sample of a new entrepreneurial form at present: the combination of OPC (One Person Company) and AI. AI tools are pushing the minimum team size for entrepreneurship to a lower level: the work that used to require a small team to complete can now be launched by one person with a set of AI tools. Almost all R&D work of Yaoshi Technology in the past two years has been completed through such a combination. For the founder, AI does not replace judgment, but amplifies it: it compresses the exploration that originally required a whole team to complete into two years of work by one person. But he also clearly knows that OPC can solve the efficiency from 0 to 1, but cannot solve the engineering problem from 1 to 100; a growing brain ultimately requires a group of people to accompany it to grow.
Yaoshi Technology is headquartered in Yangpu, Shanghai. The founder is a serial entrepreneur with 20 years of experience in the information industry. He once was responsible for the IoT system at State Grid and participated in the top-level design of major projects; in his own words, he is a person who "grows out of the system". At his request, his name is not disclosed in this article.
Building a team is the most needed part for this company to supplement. At present, the overall architecture and core mechanism of the company are independently led by the founder, and links such as neuroscience, acoustics, dynamics and benchmark testing are supplemented by external expert consultants; the R&D process is required to be documented, task-based and traceable. The founder does not avoid the key person dependency of this structure: after this round of financing, the company will reserve an option pool, introduce co-founders and core R&D members, and turn "one person's architecture" into "a group of people's engineering".
According to the plan, the funds of this round will be mainly used for intellectual property and compliance, laboratory and joint debugging verification environment, external cooperation and expert consultants, cloud resources and computing power, R&D equipment and tool chain, as well as daily operation. The company expects to use 12-18 months to complete the new round of operation period acceptance, joint debugging with the first complete machine manufacturer, voice naming learning closed loop demonstration on real robot body, complete perception-cognition-decision-action closed loop, and the first vertical scenario POC.
From the perspective of Yaoshi Technology, the robot industry does not lack more powerful bodies, but lacks a brain that can explain itself and grow in the process of interaction. Whether this brain can finally be established depends on its performance on the real robot body; but in the past two years, it has grown from an idea to a system that can run continuously.
(This article does not constitute investment advice.)