Unlimited Power: An Elite Team of Scientists Building Robot Brains | Jiazi Light‑Year

Force Unlimited, a top scientist team building brains for robots | Jiazi Light-Year

As general large models begin to enter robot decision systems, the technical boundaries of embodied intelligence are being redefined. Chen Jiayu, chief scientist at Yuanli Unlimited, believes that the robot brain still has many unresolved fundamental issues. From causal world models to 50 real‑machine continual learning tasks, from first‑person data to failure recovery in real scenarios, a Chinese embodied‑intelligence company that gathers young scientists is trying to provide its own answers.

The “Token economics” inside a ring.

Author | Zhou Yue

Edited | Chestnut

In October 2026, the embodied intelligence industry once again felt pressure from general‑purpose large models.

With the capabilities of models like GPT-6 and Astra improving, a direction that robot companies have long regarded as a core competitive advantage—task planning—is now facing re‑evaluation.

Enabling robots to understand commands, decompose tasks, and plan steps was once a key component of embodied intelligence research. Today, advances in general large models in reasoning, memory, and agent capabilities are changing the technical and economic viability of this work.

For startups betting on a robot brain, this is not an easy topic.

But INFIFORCE’s chief scientist Chen Jiayu’s reaction was quite straightforward. “If I were to make a personal prediction,GPT-6 Astra will basically eat up the entire Planner component.”Recently, Chen Jiayu was interviewed by “Jiazi Light‑Year”, and he has even decided to stop further work on some of the high‑level memory mechanisms the team previously studied.“If a general model can do this well enough, embodied‑intelligence companies won’t need to reinvent the wheel,” he said.

These remarks come from a scientist who has long studied reinforcement learning, hierarchical decision‑making, and world models, which makes them rather thought‑provoking.

Jia-Yu Chen earned his bachelor’s degree at Peking University, obtained his Ph.D. from Purdue University, and then pursued postdoctoral research at Carnegie Mellon University. He is now an assistant professor at the University of Hong Kong, head of the Agentic Intelligence Lab, and chief scientist at Infinite Force.

His research contributions span major international AI and robotics conferences such as NeurIPS, ICML, ICLR, IJCAI, and ICRA. According to the usual tech narrative, a scientist should emphasize the irreplaceability of their research direction as much as possible.

Jia-Yu Chen, however, is willing to admit that some problems may no longer be worth pursuing.

But he is not worried that robot research will lose its meaning. In his view,the stronger the general AI, the clearer the truly difficult part of embodied intelligence becomes.

A model can understand the instruction “take the clothes out of the washing machine,” break the task down methodically, and even explain why each step is performed.

But when the robotic arm actually reaches into the drum and grabs a wet, tangled bunch of clothes, all the elegant planning must face the test of the physical world.

Clothes can slip, fabric can deform, and a slight deviation in contact position can cause the next action to fail.

The robot also needs to be fast. “When executing, it must be both fast and good,” Chen Jiayu said. This happens to be the direction Yuanli Unlimited is focusing its investment on.

Among Chinese embodied intelligence startups, Yuanli Unlimited has already entered the top tier and is also one of the leading technical forces in the original research field of embodied brains. “Jiazi Light‑Year” learned that recently Zhejiang Venture Capital Group Co., Ltd. (hereinafter “Zhe Chuangtou”) completed an investment of nearly 100 million yuan in Yuanli Unlimited, just over two months after the announcement of the previous A‑round and A+ round financing of nearly 1 billion yuan.

Compared with the more familiar robot hardware companies, its attempted core competitiveness focuses on the robot’s intelligent base.

Infinite Force names this brain AtomBrain, and the team chooses the causal world model as the core technology roadmap, continuously advancing foundational models, reinforcement learning, continual learning, and real‑world data research, and connects model training with real robot deployment through the DataGrid full‑stack AI Infra.

In recent months, this company has produced a batch of highly internationally competitive research results:AtomVLA achieved a 97% average success rate on the LIBERO robot manipulation benchmark; AIM achieved a 93.1% average success rate in the RoboTwin 2.0 benchmark; SAM-3D achieved a 99.1% average success rate in the LIBERO benchmark; the robotcontinuous learning experiments progressed from 10 consecutive tasks to as many as 50 real‑world tasks.

These numbers come from different research settings and cannot be simply compared side‑by‑side, but they outline an ambitious research focus:Enabling robots to understand how actions will change the world and continuously learn new things in the real world.

It is Yuanli Unlimited founder and CEO Bai Huiyuan who brings Chen Jiayu together with a group of top researchers.

She loves “Star Wars” and has a rather personal nickname—“Star Wars Queen.” The name Yuanli Unlimited is also tied to this hobby. Today, Bai Huiyuan is turning her vision of a future intelligent world into a company with clear technological ambitions. The scientists she has recruited are studying problems that are far less romantic than those in sci‑fi movies.

For example, after a robot learns the 50th task, can it still remember the previous 49? And another example, a wet piece of clothing that falls out of a washing machine—how should we teach the robot not to make the same mistake next time?

These seemingly trivial questions may determine the future of robot brains.

1. General AI is approaching,

Where is the battlefield for robot brains?

Chen Jiayu is not keen ondefending any particular established model architecture.

Regarding the technical roadmap, he adopts an almost pragmatic attitude:Study whatever method works; if the value of a particular approach is being superseded by new technological advances, it should be adjusted promptly.

This is especially important in AI research.

In recent years, the robotics field has undergone rapid evolution across multiple technical paths—from imitation learning and vision‑language action models to world models and reinforcement learning. Researchers continuously try to answer the same question: how can robots achieve stronger generalization ability?

However, as general large models keep advancing, capabilities that once belonged to high‑level robot cognition can now be obtained externally.

Chen Jiayu judges that embodied intelligence companies will still need to research foundational models in the future, but the research focus will shift further toward execution capabilities in the physical world.

Especially in industrial scenarios.

The sequence of production tasks is usually relatively clear, and robots may not need a complex high-level planning system.What truly determines efficiency may be whether they can complete precise grasping within a limited time, handle flexible objects, and quickly adjust when contact errors occur.

In these cases, language comprehension is hard to translate directly into operational ability. A robotic arm may know where to grasp, but there is still a considerable gap before it can execute the grasp with appropriate force, speed, and trajectory.

Chen Jiayu’s research tackles precisely this gap.

  A brain must first understand the consequences of actions

After joining Yuanli Unlimited in 2025, Chen Jiayu led a team to launch a series of embodied foundational model studies.

AtomVLA is one of the earlier efforts. In robotic manipulation tasks, executing long sequences has always been a challenge. Suppose a robot needs to complete ten actions consecutively; even if each step has a 95% success rate, the probability of completing the entire task in one go is only about 60%. The more actions there are, the more errors tend to accumulate. A slight deviation in an earlier step can be amplified in subsequent operations.

AtomVLA attempts to alleviate this issue by using a hierarchical task structure and a world model to assist decision‑making.The idea is to enable the model, when generating actions, to incorporate judgments about task progress and possible outcomes, thereby reducing error accumulation in long‑sequence operations.

Research released in March 2026 showed that AtomVLA achieved a 97% average success rate on the LIBERO robot manipulation benchmark and was validated on real‑robot tasks. The work was accepted at IROS 2026.

Subsequently, Yuanli Unlimited launched research such as AIM, continuing to explore the relationship between world-state prediction and action generation.

AIM achieved a 93.1% average success rate on the RoboTwin 2.0 benchmark. For foundational robot models, these results indicate that the team already possesses a certain level of model design and training capability. However, Chen Jiayu wants to keep pushing forward.

The question he cares about is:Can robots truly understand what impact their actions will have on the physical world?

For example, picking up a cup. A vision model can identify the cup’s position, shape, and color, and an action model can generate a robotic arm trajectory based on that. But real grasping involves many more variables: where the contact point is, how much friction there is, whether the cup contains liquid, and whether the force applied by the arm will cause the object to slip.

If the first grasp fails, the robot must also determine the cause of failure before deciding how to adjust. These issues cannot be solved solely by object recognition and action generation. They involve the relationships among actions, contact, state changes, and outcomes.

Therefore, Infinite Force establishes the causal world model as the core technical route for embodied cognition.

The research goal of world models is to enable agents to learn how the environment changes; by further incorporating causal relationships, the aim is for the model to distinguish which changes are related to the agent’s own actions and to use this relationship for prediction and decision‑making.

For robots, this means they have the opportunity to evaluate the potential outcomes of different actions before executing them. This capability is especially valuable for manipulation tasks with rich contact and complex environmental changes.

The implicit causal world model that Yuanli Infinite is advancing aims to incorporate this understanding of physical relationships into embodied foundation models. Chen Jiayu explained that the team plans to finalize the architecture of the next phase integrated model around the end of October 2026, and then further scale up the training.

He also emphasized that model capability cannot be judged solely by architecture.The quality and scale of data, training methods, and subsequent real‑robot validation will all affect the ultimate performance the model can achieve. This is why Yuanli Infinite continues to invest in DataGrid.

Here, model development and data collection progress in parallel. A modeling study may generate new data requirements, and a batch of real‑world data may cause researchers to rethink how they train the model.

But even if a robot achieves a high success rate on a given task, Chen Jiayu still believes there is a more challenging problem that remains unsolved.

  Can a robot learn the 50th task and still remember the first 49?

When a robot learns a new skill, it sometimes pays an embarrassing price: forgetting old skills. In machine learning, this is called “catastrophic forgetting.”

When a model updates its parameters to learn a new task, it may disrupt abilities that were previously acquired. For systems that complete a fixed set of tasks in one shot, this issue can be avoided through independent training; for robots that need to run continuously and constantly adapt to new tasks, the cost is much higher.

Chen Jiayu wants to change this training approach.

“We don’t want to train a model for each specific task.” He hopes the robot can learn different tasks sequentially, retaining as many old skills as possible while acquiring new ones. “When it learns the 10th task, it shouldn’t forget the first nine.”

This sentence sounds like a simple request, yet it is a long‑standing challenge in the field of continual learning.

Imagine a flexible production line. In the first week, the robot assembles a certain component; in the second week, the line switches and it must learn a new operation process; in the third week, the original order returns, and the robot has to perform the first‑week task again. If the model has to be retrained every time the line switches, the cost advantage of flexible manufacturing would be severely reduced.

In May 2026, Infinite Force publicly released a continual learning study. The team built a benchmark of 10 sequential tasks involving single-arm and dual-arm manipulation, comparing how different training methods affect performance on previous tasks.

Experiments found that straightforward sequential fine‑tuning readily leads to significant forgetting, whereas retaining and replaying a portion of historical task data can effectively mitigate this issue.

Under the same computational budget as defined in the paper, the proposed method also outperformed multi‑task joint training.

But the truly interesting part lies in the follow‑up experiments. Researchers discovered that the motion amplitudes across different tasks can vary dramatically.

When a model learns a new task, if the numerical scale of the action outputs is changed, even if the action patterns of the old task are not completely lost, the actual execution may still deviate.

This is an engineering detail that is easy to overlook. For robotic arms, changes in action scale can directly affect the position, speed, and operational precision of the end effector.

Thus, the team proposed the FAN method around this issue, pre‑calibrating and maintaining a stable normalized action scale to reduce interference of new task training on old skills.

A continual learning problem ultimately gave rise to another action normalization issue.

Chen Jiayu’s team didn’t stop at the experimental results for 10 tasks. According to his description, the latest on‑robot continual learning experiments have progressed to as many as 50 tasks, involving three robotic arms. Learning a new task currently typically requires about 100 to 300 demonstration samples.

This, of course, still doesn’t prove that robots have achieved lifelong learning in the human sense.

The complexity of the 50 tasks, the learning order, the success rate on older tasks, and the training cost all need to be evaluated under specific experimental conditions.

But for a startup, being able to continuously push research problems onto real robots already says something. Especially when issues uncovered in one study quickly become the focus of the next.

Chen Jiayu seems to really enjoy this state. In his view, research has no definitive endpoint that can be declared complete.

2. The group that Bai Huiyuan recruited,

why can’t they stop?

“If we set the research direction, I can decide it by myself.” When discussing Yuanli Unlimited’s research organization method, Chen Jiayu spoke very plainly. Behind this sentence is a division of labor between Bai Huiyuan and the research team.

As a founder, Bai Huiyuan decides the long‑term direction of the company’s investment; as for the specific research problems and which technical approaches to use, she gives scientists considerable decision‑making authority.

Chen Jiayu summarizes the support she provides in four words: “human resources, compute power, capital, scenarios.”

For a startup, none of these four resources comes cheap.

Especially in foundational model research. A single experiment can consume massive compute resources, only to discover that the original hypothesis was invalid.

A model may achieve good results on public benchmarks, but after deployment on a real robot it can still fail due to data distribution shifts, hardware constraints, or environmental changes.

Researchers must accept this uncertainty. Founders need to decide how much cost they are willing to bear for it.

Bai Huiyuan chose to keep investing.She has successively recruited top talent such as Chen Jiayu and Wang Yizhou, gradually building a research team that covers world models, reinforcement learning, foundational models, data engineering, and robot control.

This is also one of the most noteworthy aspects of the Unlimited Force’s research strength—it has already formed a fairly large talent pool.

Currently, Force Unlimited has a total staff of over 150 people, with R&D personnel accounting for about 60%–70%, master’s and doctoral talent accounting for 75%, and an average team age of 27. Members’ educational backgrounds span Tsinghua University, Peking University, University of Hong Kong, Zhejiang University, Shanghai Jiao Tong University, as well as domestic and overseas institutions such as UC Berkeley, Stanford University, Purdue University, Carnegie Mellon University, etc.

An average age of 27 is the number most easily remembered about this team.

At this stage, many people have just completed their PhD training or are transitioning from academic research to industry practice.

At Infinite Force, they have already begun taking on research in various areas such as embodied foundation models, causal world models, reinforcement learning, data systems, and real robot deployment.

Educational background provides the talent foundation, but what truly determines research capability is whether these individuals can continuously pose valuable questions and produce results that stand up to scrutiny.

Chen Jiayu and Wang Yizhou are two representative figures among them.

  Chen Jiayu: A Person Who Constantly Studies “How to Learn”

If you look at Chen Jiayu’s academic background as a whole, you’ll notice an interesting pattern.

The specific problems he studies have constantly changed, butthey always revolve around how intelligent agents learn, make decisions, and control..

During his PhD at Purdue University, he was supervised by Professor Vaneet Aggarwal, a scholar in reinforcement learning and optimization, and received joint mentorship from Professor Tian Lan, an expert in multi-agent systems. He then continued related research at Carnegie Mellon University and joined the University of Hong Kong as faculty, establishing his own Agentic Intelligence Lab.

These experiences have kept him at the forefront of research in reinforcement learning, agent decision-making, and robot learning.

In 2022, his work on multi-agent skill discovery was accepted at NeurIPS.

In 2023, as first author, he published unsupervised skill discovery research at NeurIPS, exploring how agents can autonomously discover valuable behavioral skills in the absence of explicit task rewards.

In the same year, he published research on hierarchical multi-task inverse reinforcement learning at ICML, focusing on how to learn decision-making abilities with hierarchical structures from demonstration behavior.

Related robot control research also appeared at international conferences such as ICRA.

Subsequently, he continued research in offline reinforcement learning, skill discovery, and world models, with results presented at conferences like IJCAI, ICLR, and ICML.

The technical issues addressed in these studies vary.

Skill discovery concerns whether an agent can autonomously form reusable behaviors; hierarchical decision‑making concerns whether complex tasks can be organized into different levels of actions; world models further concern whether an agent can use predictions of environmental changes to improve learning and decision‑making.

Putting them together reveals the continuity in Chen Jiayu’s research trajectory.

He has been studying how agents acquire abilities from experience. Today, embodied intelligence provides a more complex object for this research.

Robots need to learn in the real world. Every action they take can be affected by friction, contact, object deformation, sensor errors, and hardware performance. This forces many algorithms that work in digital environments to undergo new scrutiny.

After joining Force Unlimited, Chen Jiayu gained access to more real robots, training data, and engineering resources, and began applying his long‑accumulated research methods to embodied brain development.

But he does not become fixated on any single technical path because of that.

Faced with the development of general large models, he is willing to voluntarily give up some existing research investments; when real‑world experiments expose issues, he is also willing to start over from what seems like a tiny action normalization problem.

This attitude is very similar to the tone he uses when discussing research.There are no grand statements, just a continual search for the next problem worth solving.

  A research model, how a research model truly works

CTO Wang Yizhou, who partners with Chen Jiayu, has a different technical background.

He graduated from the University of California, Berkeley, earning a Ph.D. in Mechanical Engineering under robotics and control scholar Professor Tomizuka.

After that, Wang Yizhou entered the autonomous driving industry, was an early core member of Baidu’s IDL autonomous driving team, and later joined NVIDIA’s DriveAV team.

Having worked in the autonomous driving field for many years, he has experienced multiple stages from algorithm development, data closed-loop, and engineering deployment to L2, L4, and mass production.

Autonomous driving has repeatedly proven one thing:the performance of a model on a test set and its ability to run stably in complex real-world environments are often separated by a large amount of engineering work.

How data is collected, how errors are discovered, how models are retrained, how new versions are validated, and how continuous improvement is achieved after deployment—these issues form the foundation for the long‑term evolution of intelligent systems.

Wang Yizhou has brought this experience to Yuanli Unlimited.

Chen Jiayu prefers to start from scientific questions, probing why robots fail to learn and how they can learn better.

Wang Yizhou focuses more on the other side: what data models need, how data is generated, how training is organized, and how algorithms run stably on real hardware and in real scenarios.

Bai Huiyuan brings these two capabilities together.

At Yuanli Unlimited, Chen Jiayu leads original research and core technical roadmaps for embodied brains, while Wang Yizhou is responsible for the full‑stack AI Infra of DataGrid and its related data, training, and engineering systems.

The division of labor between the two also explains why this company is investing in both model and data infrastructure.

For embodied intelligence, researchers often encounter an awkward situation: the data required by the model cannot be provided by existing collection methods; after painstakingly gathering data, they may find it is not suitable for training.

PrimeForce Unlimited attempts to shorten the gap between these two.

  A study on data yielded a counterintuitive result

AtomEgo is an example. The problem this research aims to solve is not complicated: how to use human first‑person operational experience to train robots?

Humans engage in massive physical interactions every day. Picking up a cup, opening a drawer, folding clothes, moving items. These actions naturally embody the relationships among vision, motion, and environmental changes.

If we could collect human first‑person operation data at low cost and large scale, robot training might not have to rely entirely on expensive real‑world demonstrations. However, the issue is that humans and robots do not share the same bodies.

The way human arms move, their joint structures, and action spaces differ markedly from those of robots.

No matter how much human data is collected, if it cannot be effectively transformed into actions that a robot can execute, it is hard to generate direct value.

Therefore, AtomEgo compared different data integration and training strategies. In experiments conducted on about 2,659 hours of data, the progressive embodied alignment approach performed best, raising the real‑robot task success rate from 32.86% to 51.43%.

Another seemingly reasonable route yielded the opposite result.

The team tried to assist model learning by predicting future visual frames, but under the related experimental settings the real‑robot success rate actually dropped to 20.71%.

This result is quite insightful. Many people think that having the model learn to predict future frames seems to help it understand how the world changes.

However, at least under AtomEgo’s experimental conditions, adding this training objective did not lead to better robot manipulation performance.

It reminds researchers:Learning more representations for the model is not always the same as enabling the robot to truly learn how to operate.

For Yuanli Unlimited, the value of this study also lies in its ability to inversely influence the design of data and training systems.

Which human operation data are worth collecting? How should data from different sources be aligned? What training order can improve robot performance?

These questions are no longer just the task of data engineering teams; they have also become part of embodied cognition research.

The same applies to haptic research. The Force Unlimited team has studied how to collect data using low‑cost, relatively lightweight haptic gloves. But building the collection device only completes half the work. The other half is more important:Can the captured haptic signals help robots learn better?

“You also have to verify that the data you collect is useful for model training.” Chen Jiayu said.

This sentence almost sums up the relationship between the internal model and data teams at Yuanli Unlimited. Data collection needs to serve model learning, and model research continuously raises new data requirements.

DataGrid thus becomes the infrastructure linking the two. From first‑person data, teleoperation, and multimodal information capture to data processing, action alignment, model training, and real‑robot validation, Yuanli Unlimited aims to build a repeatable training system.

For embodied brain companies, the significance of this framework ultimately needs to be demonstrated through the continuous improvement of model capabilities.

  “Yes, because there aren’t many people either”

Researchers conducting experiments on real robots face complexities far beyond the flowcharts in papers.

Chen Jiayu mentioned that some deployment scenarios lack stable networks or suffer from high communication latency.

Models that run in the lab may not be directly usable on site.

The team needs to consider model computational cost, inference latency, hardware performance, and sometimes decide which parts are best handled by learning models and which are more effectively addressed with traditional automation methods.

These tasks are difficult for researchers with a single disciplinary background to complete independently. Model research, data engineering, robot control, and field deployment require continuous collaboration.

Chen Jiayu leads a team of nearly 30 graduate students and postdocs. This team is also experimenting with AI agents to assist coding, algorithm discussions, and experiment debugging, with the company covering the costs of related model calls.

While AI tools can accelerate code and experiment workflows, they cannot replace the repeated physical operations of robotic arms in real environments.

When asked whether the team works closely together, Chen Jiayu answered, “Yes, because there aren’t many people.”In one sentence, she captures the reality of a research‑driven startup.

There are many research directions, but the manpower, compute resources, and time that can actually be devoted are always limited. Scientists must decide which problems are worth pursuing, which directions should be abandoned promptly, and which experiments need to be pushed to real hardware as soon as possible.

The research questions that ultimately survive often stem from the most unglamorous failures.

For example, a wet piece of clothing that fell on the floor.

3. A piece of clothing that fell on the floor,

and the future of robot brains

A wet piece of clothing that falls out of a washing machine can keep robot researchers busy for a while.

Humans would simply bend down and pick it up. But for a robot, the problem is far from that simple.

Wet clothes can become tangled, making it easy for a robotic arm to slip during grasping. The material, size, and moisture content of different garments also affect the grasp outcome.

Even if the robot knows it should retrieve the clothing, the execution may still fail. And what to do after a failure may not be covered by the existing training data.

According to Chen Jiayu, the Force Unlimited team devised a solution for garments that fall off. They placed a large laundry basket beneath the washing machine to catch the dropped clothes. Then, a person demonstrated how to retrieve the garments and continue the task. DataGrid recorded hand and wrist motions during this process and combined them with real robot data for subsequent model training.

An operation that originally meant task failure thus became new training material.

What robots need to learn has expanded from “how to correctly complete a task” to “how to keep the task going when things don’t go as expected.”

This is precisely the challenge that the physical world poses to embodied intelligence.

Exceptional situations in real environments can hardly be enumerated in advance. Clothes may fall, objects may slide, grasp positions may shift, and the robotic arm may deviate from its original trajectory due to contact errors.

If a robot can only perform the standard actions it saw during training, it will find it difficult to work autonomously over the long term; it must be capable of handling new situations.

Acquiring this capability depends on real-world data, continual learning, and post‑training.

Regarding the limitless force,the failure in the laundry scenario this time conveniently linked several research clues together.

The data system records human correction actions, the model learns new handling methods from demonstrations, and real‑world experiments verify the learning effect.

If the robot successfully acquires anomaly recovery capability, researchers still need to confirm: does it consequently forget the operations it previously learned?

This brings us back to the continual learning issue that Chen Jiayu has been focusing on.

  Real‑world training cannot rely on unlimited trial‑and‑error

Having robots learn from failure sounds natural, but each trial‑and‑error on a real robot incurs a cost.

The manipulator must operate, the hardware can wear out, and on‑site personnel are needed for maintenance. Tasks involving contact and complex object manipulation may also pose safety risks. Therefore, robots cannot, like agents in a digital environment, easily perform tens of thousands of attempts.

Chen Jiayu hopes that real‑world reinforcement learning can make meaningful progress within a few hundred attempts, without relying on tens of thousands of real interactions.The team is also researching more efficient exploration methods and universal value functions that can evaluate robot performance.

The latter is especially important. In reinforcement learning, the model needs to know which actions are good and which should be avoided. If a new evaluation mechanism must be designed for every task, the training cost becomes hard to reduce.

Research on universal value functions aims to reduce this repetitive work, allowing robots to more effectively use feedback to improve behavior.

But Chen Jiayu is not only concerned with the success rate of a single training run.

He is also concerned about another accounting issue:When one scenario has been successfully run and the robot moves to the next scenario, how much cost must be incurred again?This directly relates to the commercialization of embodied intelligence.

“Deploymentshould aim for a break‑even ROI.” Chen Jiayu said. His judgment is quite pragmatic.When a robot completes a task, the first step is to compare the total cost with the output per unit time.If the cost far exceeds that of human labor and efficiency shows no clear advantage, then even if the technology can accomplish it, it is hard to become a sustainable business model.

Another often overlooked cost is capability transfer. Suppose a robot has learned to organize a certain type of shelf. If you change stores or change product lines, do you need to collect a large amount of new data, retrain the model, or even redesign the entire action workflow?

If every new scenario has to start from scratch, the larger the deployment scale, the higher the engineering cost.

This is the business question that continuous learning, model generalization, and real‑world data closed‑loops ultimately need to answer.

Chen Jiayu hopes that a validated post‑training method can be reused across different scenarios, and that robots can operate stably over the long term.

He proposed a specific validation requirement:The system must run continuously for at least one month to test its long‑term stability.

One month doesn’t sound impressive.

But for robots that must continuously face real‑world changes, long‑term stable operation is a completely different capability from completing a single demo in a lab.

 Another achievement of Unlimited Force

Beyond research, Yuanli Unlimited has already built a highly advanced real‑world scenario foundation.

Currently, Yuanli Unlimited’s business spans more than 30 cities nationwide and over 100 real application scenarios, with cumulative order value exceeding 500 million yuan, covering e‑commerce warehousing, industrial logistics, commercial services, energy and power, among others.

These numbers and paper benchmarks are two different report cards.

The former reflects business and scenario coverage, while the latter tests the model’s capability under specific experimental conditions.

There is no natural equivalence between them. Deploying robots in real-world scenarios does not mean that all tasks have already been autonomously completed by a general embodied model; likewise, order volume does not directly demonstrate the foundational model’s level of generalization.

However, for the research team at Yuanli Unlimited, these scenarios provide another resource: a continual stream of new problems.

The real world rarely operates the way the training data does. Object positions shift, task requirements are adjusted, and environmental disturbances are hard to predict completely. Researchers can discover the limits of a model’s capabilities from these failures, then decide what data to supplement and which training methods to tweak.

What DataGrid aims to build at Infinite Force is exactly such a loop.

“Train your own model with your own data,” says Chen Jiayu.

The model is deployed into real scenarios; the problems and data from those scenarios flow back into the training system, driving the next round of improvement.

Whether this mechanism can truly accelerate the evolution speed of robots still requires long‑term validation. But it has already clarified the unlimited‑force technical choice: placing the foundation model, data generation, post‑training, and real robot deployment on the same development pipeline.

For a company focused on embodied cognition, this is harder than simply chasing a single benchmark ranking, and it is closer to long‑term competition. Because the robot will ultimately face not just fixed tasks from papers.

  Research also needs to learn accounting

Chen Jiayu does not shy away from the constraints that business operations impose on research. In his view, an embodied intelligence company must ultimately deliver tangible results and cannot rely on technical narratives for funding indefinitely.

But he is also wary of another tendency: first assuming a technology must be valuable, then searching everywhere for suitable application scenarios. That easily turns into “using a hammer to find nails.”

This statement is actually consistent with his earlier judgment when discussing general-purpose large models.

When a general model can already solve a certain class of problems, there is no need to repeat research just to prove one’s own technical approach. When real-world scenarios reveal new challenges, we should start from the problem itself and look for appropriate solutions.

For Chen Jiayu, research directions must stand up to experimental results. For Yuanli Unlimited, technological investments ultimately need to be validated in real-world scenarios.

These two tests constitute the R&D approach the company is trying to establish.

In August 2026, Force Unlimited announced the completion of nearly 1 billion yuan Series A and A+ financing; the funds will continue to be invested in embodied brain AtomBrain development, DataGrid full‑stack AI Infra upgrades, and the delivery and validation of multimodal robots in real‑world scenarios.

For a startup, this means more abundant R&D resources and higher external expectations.

Bai Huiyuan has assembled a young research team with international academic backgrounds, and Chen Jiayu and Wang Yizhou are respectively advancing the embodied brain research of Yuanli Unlimited from the perspectives of modeling and engineering systems.

AtomVLA, AIM, AtomEgo, continual learning, FAN, and the ongoing research on causal world models and real-device reinforcement learning constitute the company’s technical accumulation over the past period.

These achievements are still not enough to declare that the final answer for robot brains has emerged. Yet they are sufficient to show,Yuanli Unlimited is a leading original research force in China’s embodied brain field that cannot be ignored.

The upcoming competition will become more concrete.

• Who can enable robots to learn new tasks with less data?

• Who can make models retain old skills during continual learning?

• Who can enable robots to understand the relationship between actions and physical outcomes?

• Who can truly bring these capabilities into complex real‑world environments?

These questions are too hard to solve in a single paper, and they won’t disappear after a single model release.They require one study after another, continuous accumulation of real data, and repeated physical‑world experiments with robots.This perhaps explains Chen Jiayu’s attitude toward research.

In the interview, “Jiazi Light‑Year” asked him which research project was the most meaningful.

He didn’t pick a representative paper he had published at a top conference, nor did he list models that had already achieved good experimental results.

His answer was only one sentence:“I think it will always be the next one, the next onethat is a more meaningful study.”

Perhaps the story that the infinite Force writes for the robot’s brain will also always have a next chapter.

(Cover image and in-article illustrations source: Force Unlimited)

END.