Behind the $399 Machine Duck Sales Boom: How Shenzhen Became the 'Hardware Base' for Physical AI?

Author: Zen, PANews

A 25-centimeter tall, wobbly walking machine duck has unexpectedly become one of the most talked-about new products in the robotics industry recently.

On August 27, Pollen Robotics, a subsidiary of Hugging Face, opened pre-sales for the Microduck. Priced at $399, this robot can walk, sit, get up after falling, pick up objects, kick a ball, or glide on wheels. In addition to its core functions, developers can retrain its movements using the MuJoCo simulation environment and reinforcement learning, deploying strategies directly to the real machine.

"We didn't expect so many (orders)," said the co-founder and CEO of Hugging Face on X platform on August 31, noting that within just five days, pre-orders for the Microduck had exceeded 10,000 units. In the first six hours of sales, the order value surpassed $1 million. Due to the rapid accumulation of orders, the estimated delivery time for new orders was subsequently extended to 4 to 6 months.

The cuteness, low price, and open development environment are the main reasons why Microduck has become a hot tech item, raising an important question: How can a Microduck, equipped with 15 motors, cameras, LiDAR, dual IMUs, and local computing power, capable of running reinforcement learning strategies, be sold for just $399?

The answer lies not only in Hugging Face's software capabilities. The Microduck is shipped to the European and American markets from warehouses in France and the United States, but its country of origin is clearly marked as China. Although the specific manufacturing partner has not yet been disclosed, looking back at the production process of its previous product, Reachy Mini, a clearer Physical AI industrial chain has emerged.

From Demo to Mass Production: Shenzhen Fills the Gap in Robot Commercialization

In April 2025, Hugging Face acquired the French robotics company Pollen Robotics, officially extending its business from AI models, datasets, and development tools to physical robots. Three months later, the two parties launched the first desktop robot for general developers, Reachy Mini, which also started at a price of $399.

However, at the time of its release, the compact and open-development-oriented Reachy Mini was essentially a prototype made with a large number of 3D-printed parts, still quite distant from true consumer-grade hardware. This production method could support prototype validation and early small-batch trials, but as orders rapidly increased, the originally suitable manufacturing method for laboratory and prototype validation quickly became inadequate.

By November of last year, Reachy Mini received over 3,000 orders in just one week. The sudden increase in market demand brought a previously less urgent question to Pollen: How to transform a validated robot demo into a standardized product that can be stably produced, tested in batches, and delivered on time in thousands or even tens of thousands of units.

Ultimately, Pollen entrusted this crucial task to Seeed Studio, headquartered in Shenzhen. In less than five months, the two parties advanced their collaboration across multiple aspects, including mechanical structure, electronic systems, acoustics, manufacturing, and supply chain coordination, ultimately completing the production and shipment of 3,000 units. By the time of the Microduck's release, Pollen stated that the number of Reachy Minis in users' hands had exceeded 10,000 units.

If Seeed Studio is merely understood as a "contract factory" responsible for large-scale assembly of robots, it would completely underestimate its important role in the collaboration.

For example, Reachy Mini primarily interacts with users through voice, but the noise generated by multiple motors running simultaneously inside the robot directly interferes with the microphone array. This issue cannot be solved by simply adding an assembly step; it requires adjustments from mechanical structure and acoustic design to noise reduction algorithms. Pollen and Seeed Studio ultimately modified the product design through multiple rounds of experiments to maintain relatively stable audio pickup during the robot's movement.

This is actually closer to an engineering service. For robotics startups, creating a demo and mass-producing 10,000 products are two entirely different things.

Robots in laboratories can use expensive components, can be adjusted by engineers for assembly tolerances, can tolerate certain differences between devices, and can have human intervention at any time when problems arise. However, once entering the consumer market, robotic products must face a complete set of issues, including molds, PCBs, wiring harnesses, batteries, heat dissipation, acoustics, electromagnetic compatibility, quality control, and even after-sales maintenance.

Therefore, when robots transition from the laboratory to large-scale production, competition is no longer just about whether models and algorithms can run, but rather whether models, control, mechanics, electronics, and manufacturing can form a continuously iterative whole. This is where Shenzhen companies like Seeed Studio begin to show their value.

Behind the $399 Price Tag: Shenzhen is Not Selling Cheap Components

Returning to the popular Microduck, these cute machine ducks are also very suitable for observing this change.

According to data disclosed by Pollen, this machine is only 25 centimeters tall and weighs about 800 grams, yet it has 15 motors and is equipped with cameras, LiDAR, and two IMUs; its control system uses the Rockchip RK3566 chip, capable of running motion strategies locally at 50Hz. The official codebase also opens up MuJoCo, PPO, and Sim-to-Real training processes, allowing developers to retrain its walking, getting up, and other movements.

From the perspective of Physical AI research platforms, this hardware configuration is not luxurious. However, it is precisely this full utilization of mature hardware that provides an important foundation for reducing the price of Microduck to $399.

In the past, many robotics research platforms were expensive, not because every component was at the technological forefront, but because robotics has long been a small-batch, highly customized product: mechanical structures, control boards, motors, and sensors lack sufficient scale effects, and different components need to undergo interface development and system adaptation, while small-scale production means that mold, testing, and engineering development costs can only be spread across a limited number of products.

Microduck, on the other hand, follows a different route: it calls upon already mature, commercialized consumer electronics and robotic components as much as possible, and then creates differentiation through software, reinforcement learning capabilities, and product design.

This is also where the true value of the Chinese supply chain begins to emerge. It provides not only lower labor costs and cheaper components but also a highly mature, rapidly deployable hardware production network.

Seeed Studio has previously showcased its underlying Greater Bay Area supply chain using Reachy Mini as an example: CNC structural parts, speaker modules, heat dissipation components, battery packs, and other parts can be completed by specialized suppliers from different regions and then rapidly integrated.

For overseas robotics startups like Pollen, this means they do not need to establish a complete supply system for batteries, PCBs, molds, structural parts, etc., from scratch around a robot, but can directly tap into a manufacturing network gradually formed through decades of development in the consumer electronics industry.

Now, this industrial capability, which once served mobile phones, drones, and various smart hardware, is further migrating to the robotics industry. The "2025 Shenzhen Robot Industry Development White Paper" shows that by 2025, the output value of Shenzhen's robot industry will exceed 242 billion yuan, with a year-on-year growth of 20%; the city will produce nearly 8 million service robots, accounting for about 43% of the national total, and the output of industrial robots will reach 194,900 units, accounting for about a quarter of the national total.

More importantly, this supply chain is extending from the PCB, batteries, and structural parts of the consumer electronics era to actuators, reducers, sensors, controllers, and dexterous hands, which are key components specific to robots. For robotics companies, this means that more and more critical components that originally required cross-regional or even cross-national procurement and coordination can be completed within a highly concentrated industrial cluster.

On September 2, Shenzhen officially announced the "Shenzhen Plan for Promoting High-Quality Development of the Intelligent Robot Industry (2026-2028)", which proposed a representative goal: to promote local matching of core robot components such as sensors, actuators, reducers, and controllers, forming a "half-hour supply circle," and further streamline the complete chain from research and development, trial production to large-scale production and delivery.

How much this policy goal can ultimately be achieved still needs time to verify. However, the term "half-hour supply circle" effectively summarizes the barriers that the Shenzhen robot industry truly hopes to establish. For Physical AI, which is still in a rapid iteration phase, this capability may even be more important than pure manufacturing costs.

In the Era of Physical AI, Shenzhen is Becoming the 'Development Infrastructure' for Robots

This also explains why Shenzhen's value may be amplified again in the era of Physical AI.

In the pure software era, the core infrastructure of an AI company mainly consists of GPUs, data centers, and cloud services. After model updates, new codes and parameters can be quickly deployed to thousands or even millions of users, and product iterations largely occur at the server and software level.

Robots, however, are completely different. If engineers discover that a new motion strategy requires greater joint torque, it may mean replacing motors; changes in motor specifications may affect mechanical structure, power supply, and heat dissipation designs; structural adjustments may further alter the robot's weight and center of gravity distribution, necessitating retraining of the original motion strategy. A seemingly localized hardware modification can ultimately propagate throughout the entire system.

Therefore, the iteration of Physical AI is not merely a software upgrade, but a cycle of continuous feedback between software, hardware, and manufacturing.

In this process, the competitiveness of robotics companies does not lie in designing a product to "perfection" at once, but in how quickly they can redesign, test, modify, and produce after identifying problems, and return to the real environment for validation.

This is the real difference between "engineering capability" and traditional contract manufacturing. Traditional contract manufacturing emphasizes stable and low-cost production of products according to established designs, while the robotics industry is still in a phase of rapid change, with many products requiring frequent adjustments even after entering mass production. In this case, the true value of the manufacturing side lies in participating in and bearing this continuous iteration.

Recent reports from Reuters indicate that an increasing number of overseas entrepreneurs are seeking hardware components and prototype manufacturers in Shenzhen, and some American robotics companies are still sourcing a large number of components and hardware from China. As the global AI hardware boom heats up, Shenzhen is becoming an important destination for overseas entrepreneurs to assess China's manufacturing capabilities.

Microduck and Reachy Mini provide a concrete example of this division of labor.

In this setup, the Pollen team is responsible for the robot's product, body, control system, and reinforcement learning training stack, while Hugging Face provides the open-source AI platform and global developer ecosystem, and the Chinese supply chain, represented by Shenzhen enterprises, addresses the other end of the problem: how to transform a laboratory robot into a product that can be stably manufactured and delivered in thousands or tens of thousands of units.

Thus, Shenzhen's role is no longer just "producing robots for overseas companies." For an increasing number of robotics startups that lack the capability or necessity to build factories and complete supply chains themselves, it resembles a Physical Infrastructure, a development infrastructure for the physical world that can be directly tapped into.

In this sense, one of the changes brought about by Physical AI is that manufacturing capability is becoming part of the AI industry's infrastructure.

Beyond Supply Chain Advantages, a Smarter 'Brain' is Needed

That said, having a highly competitive supply chain in the era of Physical AI does not necessarily mean that China will achieve comprehensive leadership in the robotics industry.

A Reuters investigation into China's humanoid robot industry pointed out that while China currently possesses very strong robotic hardware and mass manufacturing capabilities, many robots still face issues of dexterity, environmental adaptability, and insufficient autonomous decision-making ability once they enter real factories. Some seemingly complex demonstrations still heavily rely on preset actions and controlled environments, and there is still a long way to go before they can autonomously respond to open scenarios.

Domestic robotics companies generally acknowledge this bottleneck. During this year's WAIC, several industry executives, including those from ZhiYuan Robotics, regarded the ability to form an effective closed loop between data, model capabilities, and hardware with real-world scenarios as one of the most critical issues for embodied intelligence.

Therefore, compared to the "brain," China's current relative advantage remains more certain in the "body," that is, whether it can produce robots, reduce costs, and rapidly complete hardware iterations.

However, the "body" and the "brain" are not entirely independent. The cheaper the robot, the more robots laboratories can purchase; the more robots there are, the more real machine experiments and data collection developers can conduct; the faster the hardware iteration, the lower the cost of new algorithms transitioning from simulation to the real world.

The Microduck, weighing only about 800 grams, illustrates this relationship very intuitively. Due to its relatively low price, a failed strategy often only results in the lightweight robot falling, with the potential hardware loss and trial-and-error costs far lower than those of large humanoid robots. Thus, $399 is not just a consumer price; it is also an AI experiment cost. This may be one of the reasons why Microduck is truly valuable and can quickly spark interest in the developer community.

What Hugging Face aims to do is not just sell more machine ducks but hopes to gradually turn the robot's behavioral capabilities into an asset similar to open-source models. After a developer trains a new motion strategy, others can download, modify, and deploy it to their own robots. Pollen has also explicitly stated its hope that behavioral strategies, training environments, and training methods in Physical AI can be shared and reproduced like software models in the future.

However, for this model to truly take hold, an important prerequisite is that robots in the real world must be affordable and easily accessible. From this perspective, re-examining Microduck reveals that it actually connects two completely different infrastructures.

One is the open-source models, data, and robot software ecosystem that Hugging Face is building; the other is the electronic, mechanical, and manufacturing supply chain formed over decades in Shenzhen and the Pearl River Delta. The former seeks to lower the development threshold for robot "intelligence," while the latter continuously reduces the manufacturing threshold for the robot "body." The two are intuitively combined in a $399 machine duck.

Of course, future global competition in robotics will not be determined by a product like Microduck. Truly universal humanoid robots will still require stronger models, larger-scale and higher-quality data, more reliable actuators, and the identification of commercial scenarios that can sustainably cover expensive hardware costs.

However, Microduck at least reveals a change that is happening: the manufacturing capabilities accumulated by Shenzhen over the past two decades around mobile phones, drones, smart homes, and IoT devices will not lose value as the industry enters the AI era. On the contrary, as AI begins to have a "body," the engineering and supply chain capabilities originally serving consumer electronics are gaining a new batch of clients—globally emerging robotics and Physical AI startups.

Thus, the future value of China's robotics industry may not only be reflected in more prominent local brands like Yushu and ZhiYuan. Beyond these complete machine manufacturers, China may also play another more hidden yet equally important role: becoming the production and engineering infrastructure for global Physical AI transitioning from code to entity.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.