Author | Huang Nan
Editor | Yuan Silai
Among a large number of Tsinghua-affiliated entrepreneurs, Gao Haichuan, founder of Qianjue Technology, stands out as a special figure. He never positions himself as an academic type. His self-definition is simple and straightforward, as he stated in an interview with Hard Krypton: “I must prioritize commercial reality first, and then focus on technology development.”
In 2019, when he was still conducting research in the Tsinghua laboratory, he keenly spotted the commercial opportunities of robot brains, but he did not rush to start a business. Instead, he kept the idea in mind for a full four years. It was not until 2023, after judging that robot intelligence had entered the window of engineering and commercial implementation, that Gao Haichuan left the university and officially founded Qianjue Technology.
This pragmatism and clear awareness have run through the subsequent growth of Qianjue.
Hard Krypton learned that Qianjue Technology has recently completed a Series A+ financing of several hundred million yuan. This round of financing was jointly participated by market-oriented institutions and industrial investors including Yuanhe Hope Capital, Xiaoguang Capital, Xinneng Venture Capital, Inno Angel Fund, Jingming Capital, and Future Frontier Venture Capital. Maple Pledge Capital has long served as a private equity financing consultant and participated in this round of investment. Since its establishment, the company has completed a total of 9 rounds of financing.
From the first day of its founding, Qianjue had a clear positioning for commercialization. At that time, the whole industry was imagining integrating ChatGPT into embodied robots, believing that large models could solve all the problems in implementation. However, Qianjue chose a path that few people took.
Gao Haichuan actually started researching cutting-edge directions very early. Back in 2017, when the core members of the team were still in the Tsinghua laboratory, they had locked in the technical main line of the predictive world model, and achieved excellent results in international competitions.
But Qianjue did not publicize such stories in the early stage. When everyone was indulging in the future of general-purpose humanoid robots, Gao Haichuan said: “There is no such thing as the ‘ChatGPT moment’ in the robotics industry. There will not be a single point of technological breakthrough that brings an inflection point of instantaneous explosion across the entire track.”
He is not pessimistic or bearish about the industry, but has recognized the reality in actual implementation. Qianjue’s first batch of orders was not for humanoid robots. Instead, it chose to cut into products and scenarios with clear existing demands such as catering, cleaning, and hotel service robots.
Gao Haichuan clearly saw the problems that would be encountered in implementation, and he did not think that the problem of data collection could be solved in the short term. Therefore, Qianjue operates two lines in parallel internally: the R&D team follows the main line of the predictive world model, based on the polynomial representation architecture, to explore all possible segmented algorithm routes; the implementation team plunges into scenarios such as hotels, retail, and home services, working with clients one by one to optimize in real applications.
Robots equipped with Qianjue’s world model perform autonomous delivery in a coffee shop (Source: Enterprise)
“Clients don’t care about your technical architecture or algorithm models, they only care whether this thing can help them tidy up the scattered slippers in hotel guest rooms,” Gao Haichuan told Hard Krypton.
In the eyes of Zhang Tianren, CTO of the company, Gao Haichuan is an “atypical Tsinghua alumnus”. Although he has long-term experience in technical research, he does not take the number of papers or technical concepts as the only evaluation criteria for entrepreneurship. Gao Haichuan rarely intervenes in specific technical solutions, but focuses his energy on business directions, discussing with the team what to do and what not to do, while clarifying the long-term development path of the company.
He is also rarely influenced by external trends, and has a clear judgment on hot topics. For example, regarding brain companies developing their own robot bodies, he believes that “The high possibility behind this move is to raise financing. This in itself is a microcosm of the industry’s impetuosity, which is not based on technical logic, but on capital logic.”
After the embodied intelligence track has attracted massive capital inflows, it is now time to deliver tangible progress. But obviously, many companies still cannot produce practical results other than beautiful PPTs and exhibition hall demos.
Qianjue does not seem to have such anxiety. Gao Haichuan never talks about abstract concepts. “No matter how advanced the technology is, if it cannot be implemented in real scenarios and gain client recognition, it has no practical value,” said Gao Haichuan.
The following is the transcript of the conversation between Hard Krypton and Gao Haichuan, founder and CEO of Qianjue Technology, the content has been edited:
Investors no longer pay for cool demos
Hard Krypton: From the perspective of front-line industrial practitioners, what is the most intuitive and fundamental industry change you have observed in the embodied intelligence track this year compared to previous years?
Gao Haichuan: At the current stage of industry development, the development speed is often the fastest, and it has moved away from the value definition method of the academic circle.
In the past, people focused on one-off demonstrations, demos, and ranking lists; now they pay more attention to real implementation indicators, such as task success rate, continuous operation time, anomaly recovery capability, deployment cost, and data iteration speed. The evaluation criteria of investors have also changed, and they will no longer invest just because of a cool demo.
This change is taking place simultaneously at home and abroad. We can also see that the key words of industry evaluation criteria are converging around several dimensions, and concepts such as long-tail feedback closed loop and system engineering are being mentioned more and more frequently.
Although there is no unified architecture that can adapt to all scenarios, a consensus is taking shape. As a product, this is no longer a one-time delivered SaaS system, but a complete set of system engineering, which relies on the model system and real user feedback data to continuously iterate and co-evolve in the actual implementation and operation process.
Hard Krypton: In the past year, technical hotspots have shifted rapidly from end-to-end, VLA to world models. How do you view the market’s attitude towards new routes or concepts?
Gao Haichuan: In fact, the endless emergence of various routes is a common phenomenon in the current AI field, not only in embodied intelligence, but also in multi-modal models and language models. But the key point is that the technical foundation has not changed. At the sub-level, technical routes are not mutually exclusive. For the system framework, usually only one decoder needs to be added, and the backbone network and feature representation can be reused.
To judge whether a route is feasible, we cannot look at short-term popularity. Any segmented route has its strengths and shortcomings at a specific stage. So “new” does not mean “good”. The real test standard is the market, which can be measured by several basic indicators, such as data efficiency, sample efficiency, iteration cost, system stability during long-term operation, and generalization ability. These are the factors that determine how far a technical route can go.
Hard Krypton: Qianjue has been working on predictive world models since the first day of its establishment. At what time point did you decide to take this route?
Gao Haichuan: The domestic boom of predictive world models started after Yann LeCun founded AMI Labs, when a group of teams that had long been engaged in related research and had not received sufficient market attention came into the spotlight. Qianjue is one of them.
In fact, when the first paper in the World Model field was published in 2018, our team had already started research on Model-based Control based on the VIZDOOM video game competition, and achieved excellent results in international competitions.
At that time, we saw a long-term direction with extremely high certainty. Therefore, when the world model became a hot topic this year, Qianjue had accumulated nearly ten years of experience at the technical level.
Quantifying and Optimizing Simplicity via Polynomial Representations, which proposes to use polynomial representation to quantify and optimize the “simplicity” of neural networks (Source: Enterprise)
Hard Krypton: Before the concept of world models became popular, what was the industry’s real understanding of it? Why didn’t Qianjue choose the more mainstream VLA route in the early stage?
Gao Haichuan: Qianjue was founded in 2023, when the industry generally believed that connecting ChatGPT to humanoid robots could solve all problems. But this judgment obviously underestimated the scarcity of data. The collection cost of real robot data is extremely high, with very high time cost and damage cost. This means that the algorithm must have extremely high sample efficiency to cross the threshold of commercialization.
VLA and world models are not completely opposing technical routes. VLA has its own advantages in language understanding and direct action generation. But if you only rely on end-to-end strategy learning, you may still face high data demand and generalization challenges when facing data scarcity, environmental changes, and out-of-distribution tasks. Therefore, Qianjue was very clear from the beginning that it would not take the Model-free route, but choose the predictive world model, which had not even become the industry consensus in the embodied brain field at that time.
Under the condition of limited data, we hope to improve the data utilization efficiency of the model, cross-scene generalization ability, and continuous decision-making ability in dynamic environments.
Another point is also crucial. The dynamic random interference in open scenarios is exactly the strength of the Transformer architecture; robots cannot only generate one action based on the current input, but also need to continuously update their judgment of the environment based on new observations. This requires the model to not only have the modeling ability of multi-modal information and time series relationships, but also strong cross-scene generalization ability.
We chose the predictive world model because we hope to learn the state representation and change rules related to tasks, so that robots can predict the possible results of actions, continuously adjust decisions and plans in dynamic environments, and reduce dependence on fixed scenarios and existing data distributions. These factors together support our technical judgment.
Developing robot bodies should not be used as a financing tool
Hard Krypton: As a provider of robot intelligent infrastructure, how do you understand the real demands of complete machine manufacturers when choosing external intelligent bases? What are the common concerns of clients?
Gao Haichuan: The biggest concern of complete machine manufacturers is actually niche conflict. For a manufacturer that already has B-end clients and its own robot body brand, if it finds that the supplier also has a robot body brand, even if there is only a potential possibility of competition, the cooperation will most likely not move forward. This is not a technical problem, but a trust problem.
We hope to adapt to different forms of robots, reduce the repeated R&D and adaptation costs of robot body manufacturers in terms of intelligent capabilities, and help partners complete product implementation faster, instead of competing with partners for robot body brands or end clients.
As for developing robot bodies, it may have special financing value in China, and even become a “revival card” at certain stages. This in itself is a microcosm of the industry’s impetuosity, which is not based on technical logic, but on capital logic.
Hard Krypton: The contradiction of robots is that the demo indicators in the laboratory are very beautiful, but once migrated to the real physical world, the performance may drop sharply. To achieve large-scale commercial replication, what is the biggest bottleneck: algorithm, engineering adaptation, or data closed loop in real physical scenarios?
Gao Haichuan: After years of iteration, the capability reserve of algorithms has actually been relatively sufficient. The real bottleneck at the moment is that the data is not really in place. Among the three elements of computing power, algorithm and data, computing power and algorithm are basically ready, and the data supply is the factor that really slows down the progress.
The scarcity of robot data is a unique dilemma of this track. The real data of the physical world can only be actively collected by physical robot entities, which is completely different from large language models. Language models can inherit the massive public text accumulated over decades on the Internet; but the robotics industry has not experienced a historical stage of large-scale manual operation of robots to accumulate large-scale physical interaction datasets.
Without this innate accumulation, the industry used to pin more hopes on algorithms and computing power in the past, but the data short board cannot be bypassed, and new solutions must be found.
This is the reality we must recognize, There is no such thing as the ‘ChatGPT moment’ in the robotics industry. There will not be a single point of technological breakthrough that brings an inflection point of instantaneous explosion across the entire track. The forward path of the industry can only be driven by real scenarios one by one, and complete market penetration step by step.
Robots equipped with Qianjue’s world model autonomously perform table cleaning tasks (Source: Enterprise)
Hard Krypton: How to judge which robot capabilities can be reused across different bodies, and which capabilities are inherently unable to be migrated across different bodies?
Gao Haichuan: At present, the industry has basically reached a consensus that task planning and the understanding of the interaction logic of objective objects can be reused across bodies. Whether it is a quadruped robot with a robotic arm, or the robotic right hand of a humanoid robot, the cognition of how the cup will displace and how its state will be changed by external force behind the action of grabbing a cup is interlinked. It answers the question of “how will the cup be operated”, not “how should my body operate the cup”.
The part of “how should I operate the cup”, which corresponds to the control of the actuator body, is deeply bound to the hardware form. Whether it is a quadruped robot, drone or humanoid robot, the motion control logic of the body is completely different, and this part cannot be directly migrated across bodies.
The upper limit of the capability of the “one brain with multiple forms” cross-body reuse model is determined by the data base. Even if the model architecture is designed to be advanced enough, once it lacks the support of data of corresponding distribution, the task success rate in real scenarios will drop significantly. Qianjue hopes to precipitate common capabilities such as task understanding, environment cognition, and decision planning, combine and adjust them according to the capability boundaries of different bodies, so as to improve reuse efficiency while retaining targeted adaptation for specific execution systems.
Hard Krypton: From the industrial side, B-end clients do not pay attention to the differences between technical routes such as VLA and world models. What are the core evaluation criteria and real client demands for the commercial implementation of embodied intelligence at the current stage?
Gao Haichuan: Industrial clients do not care at all whether you use VLA or a world model, what they care about is: can this thing allow me to hire one less employee? Can it reduce the picking error rate in the warehouse? The implementation of embodied intelligence is essentially solving these trivial and specific problems.
Qianjue Predictive World Model (Source: Enterprise)
For example, the