Written by | Yuan Yi, Edited by | Wu Xianzhi
At the 2026 WRC, embodied AI shows two distinct trends: one emphasizes strong data and a powerful AI brain; the other focuses on a robust embodiment and integrated machines. It seems that core component manufacturers face the risk of being replaced, merely called upon, or acquired as long as they do not build the complete machines or the AI brains.
The industry is permeated with a sense of convergence, yet the data, training, and models for the AI brain show no signs of converging. On the embodiment side, whether it is wheeled bases versus half-body designs, grippers versus dexterous hands, or even targeting B2B versus B2C markets, there is not even a faint scent of certainty.
For instance, while the technical routes for dexterous hands have not yet converged, the questions from customers have already changed. At the WRC venue, Linker Hand simultaneously showcased products based on different technical routes, including linkage, direct drive, and tendon-driven mechanisms. During the exchange with co-founder Zhang Yanbo, a recurring shift emerged: compared to merely focusing on degrees of freedom (DOF), payload, and weight, an increasing number of customers are now starting from specific tasks, inquiring about success rates, lifespan, efficiency, and deployment costs.
As customers begin to care about "whether this can actually be done in the long term," the competition faced by dexterous hand manufacturers is shifting from hardware specifications to scenario-based delivery.
Photon Planet brought these questions to have an in-depth communication with Zhang Yanbo, co-founder of Linker Hand. He responded to them one by one with practical experiences, industry background, and the current progress of Linker Hand.
An interlude occurred during the communication: a purchaser claiming to be from an Apple factory stated that he really did not want to wait for others to finish their conversations and just wanted to ask one question: "I need to procure equipment. I asked several staff members, but none of them could make a decision. They told me to ask you. Now, find me someone who can make decisions just like you to introduce the products."
Hearing that there was a major client, Zhang Yanbo did not terminate the conversation. Instead, he gently called over the liaison colleague and returned to the conversation once again.
Photon Planet has edited the content of this communication without altering the original meaning of the dialogue. The following is the transcript of the conversation:
Photon Planet: Integrated machine manufacturers are all pushing into dexterous hands. As a leading dexterous hand manufacturer, what is the irreplaceability of Linker Hand?
Zhang Yanbo: The biggest difference is that besides hardware, we are also developing models, especially the capability to manipulate objects made of soft materials. Another point is that in the current global retail scenarios, technical solutions have not converged. We are the only manufacturer in the world that has independently developed all three technical routes (linkage, tendon-driven, and direct drive) from the bottom up.
As the industry continues to develop, there is actually no separation between the embodiment and the AI brain. It only depends on who can solve problems in the scenarios; whoever does that is the embodiment. Whoever can train the model and solve the flexible assembly problem will get orders from customers with such needs. Therefore, at the delivery end, there is no fixed form of the embodiment. Suppose a manufacturer claims that the neck has a very high degree of freedom, but the customer needs to calculate the cost-performance ratio. In the actual delivery scenario, the most important thing is whether the degree of freedom of the neck has value for the delivery.
Photon Planet: Have there been any attempts to integrate the three routes?
Zhang Yanbo: No, the necessity for integration is insufficient. Humans have different occupational divisions of labor: athletes' hands pursue strength, doctors' hands pursue precision, and couriers' hands pursue durability. Human hands have different characteristics according to occupational divisions to meet the scenario needs of the professions. Therefore, different routes are inherently suitable for certain scenarios.
Different application scenarios require different hands. Currently, the industry evaluates the good or bad of dexterous hands from different perspectives. If it is more friendly to training and model building, it will definitely have to sacrifice the characteristics of another solution in specific scenarios. Therefore, we need to evaluate whether to accommodate the model capabilities or to accommodate future application scenarios.
Photon Planet: Tendon-driven mechanisms are one of the star product lines of Linker Hand. However, once used for a long time, creep and wear will occur. Are there currently any solutions or optimization plans for this? Is it possible for the tendon cables to become a replaceable consumable, just like changing the strings of a violin?
Zhang Yanbo: This is a material and structural issue. All we can say is that everyone is working hard on it. For currently cost-effective materials, there is no perfect solution yet, so we can only iterate continuously.
Playing at home, changing strings by yourself is harmless, but if you participate in a music competition, you still need to find someone to tune the scale. Applied to the robotics industry, if you just grab an apple and make a demo for everyone, changing it by yourself is OK. But if you need to perform very fine operations, such as the precision of replacing a patient's teeth, where not even 0.1 millimeter can be off, if it is not well verified at this time, once there is a deviation in precision, it will conflict with the data you trained before, and the model will fail.
Photon Planet: Since there is no obvious path convergence, will pursuing all three solutions bring greater challenges?
Zhang Yanbo: Yes, all three solutions have their own advantages and disadvantages; otherwise, there would be no divergence in technical routes. If any one had an obvious advantage, everyone would immediately swarm to it, and there would be no controversy.
Photon Planet: Among the current customers of dexterous hands, have there been structural changes in embodiment manufacturers and B2B customers?
Zhang Yanbo: This is a gradually changing process. The industry has been accelerating for only two or three years in total. Last year, we basically sold to integrated machine manufacturers, universities, and AI brain companies. By this year, 70% of the companies are demonstrating capabilities in various scenarios, and everyone is working on solutions for different scenarios.
When making solutions, we do not rely on humanoid forms. We have developed dual arms and dual hands ourselves and turned them into independent solutions. Now, on the one hand, we provide core component solutions for downstream customers; on the other hand, we provide our own scenario-based solutions for other customers. Both markets are incremental markets, only the proportion will not be as drastically different as before.
Photon Planet: In the component segment, Linker Hand can adapt to other manufacturers through universal interfaces, which is essentially an engineering issue. Then, in the AI brain segment, how does Linker Hand ensure compatibility with other manufacturers?
Zhang Yanbo: On the one hand, we customize dexterous hands that meet scenario needs for embodied AI companies. On the other hand, during the delivery process, we can directly deliver the dual arms and dual hands + vision + tactile solutions according to the scenarios. Once entering the scenarios, it depends on customer needs; what customers definitely want is the most cost-effective solution.
Photon Planet: Dexterous hands are very suitable for fine movements, but they are also prone to damage during work. For factories that emphasize ROI, this will bring high cost pressure. Does this constrain the applicable scenarios for dexterous hands?
Zhang Yanbo: At the 2024 WRC, the opening and closing cycle count for products from any company focusing on dexterous hands was 50,000 times. At the 2025 WRC, the industry average level was between 300,000 and 500,000 times. At this year's WRC, we have achieved 1.75 million times. From these three time nodes, we must dynamically look at the issues of product lifespan and cost. We cannot assume that it cannot meet industrial needs just because of a few tens of thousands of times, ignoring the mutual iteration of hardware and algorithms, continuously fine-tuning the business model, and ultimately achieving a qualitative change from quantitative changes. One step every year, accumulating day by day, we slowly enter into the scenarios.
Even after entering the scenarios, customers will not immediately think it is easy to use and give you 10,000 orders. The implementation is gradual. At the beginning, they give 20 orders, and they will continue to add orders only if it is done well. At the same time, our own hardware and algorithms are also continuously iterating, completing the implementation in this two-way gradual process.
Just like autonomous driving, everyone knows the ultimate goal is L4, but we cannot say that because we have not reached L4 now, we cannot sell cars below L4. If so, the entire industry would have nothing to talk about.
Photon Planet: In this year's WRC venue, many manufacturers are demonstrating logistics implementation. Linker Hand, targeting high-value fine scenarios, precisely forms a clear differentiated competition with the industry?
Zhang Yanbo: Of course.
Photon Planet: Does high-value fine scenarios mean that data value is equally important?
Zhang Yanbo: Yes, why do we want to do product assembly solutions? Because assembling joint modules is already a relatively fine component, we can build a complete data loop from the full set of modules. From assembling a robotic arm to a dexterous hand, we have the full set of data ourselves.
If generalized to assembling Apple phones—by the way, the person just now is a customer from an Apple factory. Well, since we just entered the scenario, I don't have the full set of data for Apple. We can only use the data we currently have to train our capabilities first. When the capability reaches a certain level, we can find companies that are one-tenth or one-twentieth the size of Apple, provide some solutions, and they might accept us. Then we accumulate data bit by bit, improve scenario application capabilities, and iterate step by step.
Photon Planet: How does Linker Hand build sufficiently stable, abundant, and high-value data assets from the core segment of dexterous hands?
Zhang Yanbo: By providing customers with sufficiently good solutions and expanding into sufficiently many scenarios, we can naturally build up the scenario data. It is unrealistic to say that if we sell one dexterous hand, we have the solution for one hand. It is possible that we sell 100 hands and have the data for 10 hands, because these users are willing to share data with us to help them solve more things better.
Photon Planet: What are the current areas of advantage for Linker Hand? Is it more challenging for large companies to establish feedback loops?
Zhang Yanbo: As you just saw, retail manufacturers have the most motivation. On the one hand, they have extremely strong cost pressure and are eager to introduce dexterous hands to solve the time and regional coverage gaps of human labor. On the other hand, their tolerance for data itself is much higher. However, large companies like Apple, especially in high-value fields, have very high compliance and security restrictions.
Photon Planet: Lacking feedback, how to transition through the validation period and break through large companies like Apple with better solutions?
Zhang Yanbo: No matter how large a company is, it exists within the industrial chain. We can start with other companies in the industrial chain, polish the products through cooperation, and then gradually enter the factories of the chain leaders.