HomeArticle

Stellar Intelligent Computing has closed a multi-million-yuan strategic financing round to develop the "brain cognitive black box" for embodied intelligence.

林间明月2026-11-19 11:33
Stellar Intelligent Computing has secured financing to develop an AI cognitive scheduling layer that empowers embodied intelligence

Recently, Hefei SentiMind Artificial Intelligence Technology Co., Ltd. (hereinafter referred to as SentiMind) has completed millions of yuan in strategic financing (Seed+ round), with the investor being Beijing Jingshi Zhongjiao Group. The funds will be mainly invested in the evaluation of cognitive learning ability of AI cognitive capabilities in education scenarios, improve the brain power data collection and model training covering the whole process of human thinking, memory and decision-making, and accelerate the research from carbon-based life to silicon-based intelligence.

Founded in 2025, SentiMind focuses on the R&D of multimodal foundational models for cognitive science, and "moves" the research paradigm of cognitive science to AI: it collects human behavioral data through gamified cognitive tasks, then uses multimodal models to understand, predict and train cognitive capabilities. It has now built a closed-loop system covering the full lifecycle of "cognitive assessment - training - tracking". The data covers multiple age groups from preschool to adulthood, recording the whole process of students completing assessment tasks and millisecond-level behavioral temporal features, including but not limited to: reaction time (RT), click/slide trajectory, accuracy rate, error patterns (false alarm/miss), strategy switching frequency, attention decay curve, task tolerance and Trial Snapshots in classic cognitive paradigms such as Stroop, N-back, task switching and so on. The collection and annotation of such whole-process, millisecond-level data features high threshold, long cycle and great difficulty in standardization, and large-scale high-quality data is rare in the industry. From SentiMind's perspective, the scale, dimension and collection standard of data itself is a barrier, and also the foundation for continuous iteration of cognitive models.

Large models are evolving from "being capable of fluent speaking and writing" to "being reliable and usable". From SentiMind's point of view, the deficiency of "cognitive capabilities" such as memory, reasoning and decision-making is an important root cause of problems including hallucination and amnesia; the assessment paradigms and data accumulated by cognitive science can exactly fill this gap. Thus it forms a path of "serving humans at one end and serving AI at the other": for humans, it serves scenarios including children's learning difficulty screening, talent discovery, adult career matching and elderly cognitive decline; for AI, it provides scarce cognitive data and mechanism inspiration for the next generation of large models.

The "lopsided performance" dilemma of large models: sufficient computing power but lack of mental awareness

When everyone is competing to optimize large model parameters and motion control of humanoid robots, AI in the real physical world is still like a "student with unbalanced subject performance" — it can recognize images and carry out smooth conversations, but does not understand what "fatigue" is, does not know when it will be "overloaded", and even lacks respect for the common sense of the physical world. To fill this missing part, SentiMind has chosen a differentiated path: starting from brain science, to build a "cognitive scheduling black box" for AI that is independent of large models.

At present, the mainstream route of the embodied intelligence and humanoid robot industry is to deploy end-side VLA (Vision-Language-Action) models. Although they have solved the mapping problem from "perception to action", they generally have a fatal weakness: the lack of an independent cognitive scheduling layer.

"The current robot brain only knows how to generate outputs, but does not know how to exercise restraint," said Zhang Zhong, founder of SentiMind. In actual implementation, end-side models often face three major pain points: first, when dealing with long-sequence complex tasks, the system is very likely to crash or generate disordered instructions due to working memory overload; second, they only understand text probability, lack "inhibitory control" of the physical world, and often output unexecutable dangerous instructions such as "grasp fragile items at high speed" (namely physical hallucination); third, forcibly running large models with high parameter counts locally leads to persistently high BOM cost and power consumption of computing power boards.

To solve these pain points, simply optimizing the large model itself has reached a dead end. The solution proposed by SentiMind is "layering" — they do not build omniscient and omnipotent general large models, nor do they manufacture robot limbs, but build an end-side cognitive scheduling black box (Senti-OS).

This black box is centered on a lightweight small cognitive model of 1-3B parameters, acting as the "prefrontal cortex" of the robot. When the large model outputs instructions, the black box will conduct cognitive load assessment and physical feasibility verification. It can teach the robot to actively downgrade tasks when computing power is insufficient, and forcibly intercept dangerous actions, so that the robot can not only "move", but also "move properly" and "move intelligently".

From human cognition to silicon-based intelligence: hard-to-replicate data flywheel

SentiMind's confidence to build the "cognitive black box" comes from its extremely scarce data assets behind it.

In the AI industry, text corpora are already overflowing, but "real human dynamic cognitive temporal data" is extremely scarce. Most traditional robot manufacturers rely on simulation synthetic data, which cannot simulate the real human rules of attention decay and error making.

Through early ToB/ToG commercial implementation (such as cognitive screening in public schools and assessment in education and training institutions), SentiMind has accumulated more than 3 million real desensitized cognitive behavior data, a library of more than 500 gamified cognitive tasks, and the V1.0 large cognitive behavior model.

"Using the rule of real human 'energy depletion' to train the scheduling model of robots to 'act according to their capabilities' cannot be replicated by pure robot data," explained founder Zhang Zhong.

At present, SentiMind's underlying ToB business has achieved closed-loop operation in multiple public schools and leading education and training institutions in Hefei. These commercial scenarios not only provide the company with healthy cash flow, but also become an "arsenal" that continuously produces scarce data, feeding back the training and iteration of the end-side small cognitive model.

The "iron triangle" of top-tier academia, engineering and business

To open up the extremely long and complex chain of "brain science - large model - end-side hardware", a team with strong cross-border capabilities is indispensable. At present, SentiMind has built an "iron triangle" composed of scientists, engineers and business operators:

· Academic side: Supported by relevant professors from the State Key Laboratory of Cognitive Neuroscience and Learning of Tsinghua University and Beijing Normal University, to ensure the absolute scientific validity of the underlying cognitive paradigms;

· Engineering side: Lv Gang, CTO, is the former CTO of BTG Group, with 20 years of experience in complex system architecture, responsible for "slimming down" and quantifying massive brain science algorithms and adapting them to end-side chips;

· Business side: Mei Jian, COO, is an MBA graduate of Tsinghua University, who has managed well-known brands such as 8848 Phones and Xiao Guan Tea, responsible for the commercialization of underlying models and large-scale channel expansion.

Conclusion

At present, SentiMind is at a critical node of leaping from "software algorithm" to "end-side chip adaptation". The company has officially launched a new round of angel round financing. The funds will be mainly used to recruit embedded AI architects, complete the quantitative packaging of the 1-3B small cognitive model, and cooperate with leading robot manufacturers to promote real machine verification.

In the trillion-level embodied intelligence track, large models are the muscles, VLA is the nerve, and SentiMind aims to become the "prefrontal cortex of the brain" that determines the upper limit of the robot's mental awareness.