In four seconds, you can sell a hugging face, a recently launched hugging face that sells for $399.
On the outside, this "dwarf" looks numb, not only walks, sits, bends, but also slides on dry ice, rubs things with his mouth, even gets knocked down and can stand on his own.
In addition, it was not designed to be a mouth-breaking AI assistant, but only to produce some strange, closer to the smallest animals, "spoken phrases".
This is a toy-like robot, with pre-orders exceeding $1 million in a few hours, and 24-hour sales exceeding $2.6 million (about 5,000 units). Due to orders being too large, the official announcement has delayed the delivery of the new orders to 4-6 months, with the first products expected to be delivered after Christmas.
This makes it one of the most sought-after consumer-grade robots in the field of physical intelligence in 2026, and the media has called it "the golden age of physical intelligence".
Thomas Wolf, co-founder of Hugging Face, said that Microduck's ultimate cumulative sales target was 20,000 units. Because the best-selling robot product in the robotics industry is around 20,000 units, Microduck has the potential to break this record.
Microduck is so flammable that it doesn't have any functionality.
The official simulation can be accessed in the browser.
It is fully open-source in terms of software and algorithmic ecology. Developers can use reinforcement learning algorithms (such as PPO algorithms) in MuJoCo simulators to train the various new behaviors and strategies of machine learning, and then export the trained neural network strategy to ONNX format, deployed directly to Microduck's onboard chips, to allow the machine learning to reproduce new actions in real life.
Microduck's open source warehouse, the number of stars has surpassed 6,000.
Moreover, it's difficult for developers to be free to write and modify the motion control and interactive logic of the machine gun, and to share it with others in the Hugging Face community. Many people who pay attention to it think, "I know it's useless, but I just want to buy one".
Next, let's look at what games the developers are currently developing.
It hasn't been delivered yet, it's been played.
The real hardware of Microduck has not yet been delivered, and community development has already begun.
This is due to the fact that simulators, robot models and training codes are released at the same time as the official release of the product. Even without spending $399, developers can adopt a virtual mouse in a browser or local computer first.
Some trained it to roll its head, stand upside down, bow and dance; others designed obstacle races for it to find a place to fall between a pile of rigid bodies. Developers trained a "staircase-climbing" strategy: facing a 2.5 cm high ladder, the "neck" first steps up the ladder, then stops with its head and keeps its balance.
The following is a link to the strategy:
The author specifically writes on the page that the strategy is currently only tested in a simulator and has not yet been verified on a real machine.
Similar strategies are also useful for mouth-throwing:
The following is a link to the strategy:
Another project moved the Microduck into AR. Opening the camera to scan the ground, you could put a 25cm high virtual box in the room in real proportions and control it with a mobile phone rack to walk, play, roll, and make you feel like you're buying it home in advance.
Address of the project:
Some felt that the official sound was insufficient, so Microquack was developed to create a new "voicemail" using high-pitched curves, ringtones, ringtones and noise.
Address of the project:
Simulators are also connected to large language models and MCPs, allowing users to command in natural language, walk, sit, play, etc., using a combination of models.
Address of the project:https://github.com/rokbenko/quackd
More radical developers have begun to consider adding it to the Home Assistant. Ideally, the mouse can navigate the house, perceive the environment with cameras and deep sensors, and turn on the lights and play music through the smart home interface.
In South Korea, the machine is responsible for the muscles.
China is responsible for chip + rework
Microduck was developed by the French robotics team Pollen Robotics under Hugging Face.
Previously, the company had launched its flagship "person-to-person interaction" desktop robot, Reachy Mini, with cumulative sales of over 10,000 units.
This time, they shifted the direction from "will interact" to "will act" behind all sorts of creatures, trying to figure out how AI operates in the physical world.
So don't look at this piece of shit as small, but the configuration is called the "Five Dirty Clubs".
The whole machine is only 25 centimeters tall, weighs less than 800 g, and comes to the table with no pressure.
The core model is the XL330, produced in South Korea by ROBOTIS and distributed over both legs, neck and head, producing a 14-dimensional control signal.
With this set of "small body boards, big joints", the Microduck can not only swing, walk, get up, play balls, but even slide directly onto the skates.
It is small, but it is also a great advantage of reinforcement learning. After all, the robotics movement is essentially constantly trying to make mistakes.
In some ways, training with physical intelligence in a robotics lab is becoming something that can be done at a desk.
It doesn't just sell chicken.
The clutch is controlled by an independent clutch, while part-time lightweight clutches allow Microduck to have the most basic Manipulation capabilities.
However, limited in body size, it can only hold about 100 g of objects, the weight of a piece of soap, or small objects such as a pencil or pencil.
Microduck structural analysis
Interestingly, after the release of Microduck, even Yuke's successfully marketed its own servo motor.
However, Pollen said that while the current model is more expensive, it is not directly compatible with Microduck.
In order to perceive the surroundings, the body is equipped with a front-facing camera, microphone, speaker, laser radar, and two motion sensors, which basically cover visual, auditory, distance perception and self-motion detection.
And the one that really carries the brain is a Chinese core.
The RK3566, which is powered by a Rockchip, integrates a four-core Cortex-A55 CPU with about 0.8 TOPS NPU AI.
Originally widely used in smart terminals, development boards, and edge computing devices, the SoC is now directly embedded in the body of a number of robots, and is used to reason on the edges of 50Hz real-time control rings and reinforcement learning strategies.
It also comes with 1GB of memory plus 32GB of storage.
In addition to lowering brain costs through China's already highly mature consumer electronics and edge AI chip supply chains, the Chinese manufacturer Seeed Studio, a well-known open-source hardware manufacturer, is responsible for delivering technology.
In terms of batteries, it uses a removable NP-F550 camera battery, with a capacity of about 2600 mAh. This was a standardised battery that was very common in photography and video equipment, cheap, affordable, and can be replaced at any time.
It lasts about an hour, depending on what the rabbit is doing, so it may need to be rested before you rest.
For example, the $39 package included two rollers, a ball, a laser indicator and an NFC label. Pollen also sold additional batteries and development supplies separately.
Real focus: reinforcement learning + Sim2Real, fully open source
If the Microduck can only be remotely controlled for two steps, it is at most a high-end electronic toy.
Pollen has provided Microduck with a full suite of open source software, including: robot control, simulation, reinforcement learning, and Sim2Real deployment tools.
The corresponding code has been opened in sync.
Microduck's main warehouse is responsible for the SDK and robot kernel software, while another Microduck RL project is dedicated to reinforcement learning training and Sim2Real.
Developers can first train virtual monkeys repeatedly in a simulated environment, and then deploy the same set of strategies to the real Microduck.
Pollen also discovered a "non-scientific" law: The greater the number of microducks, the greater the fun.
One can walk, two can race, three can already start folding around a football, four can even race directly.
But for developers, what's really interesting about multi-robot is that it can naturally become a low-cost Multi-Robot System experimentation platform.
Public reproduction is about to be released
Over the past two days, Microduck's replica projects have gradually increased, such as microduck-replica and Fanduck (personal 1:1 physical replica record).
The former comes from the officially released MJCF simulation model + 47 STLs, which re-introduces full assembly relationships, explosive charts, and CAD-enabled assemblies (precision up to 0.1mm).
Address of the project:https://github.com/fanhao375/microduck-replica
The latter is a personal project that is actually underway, with public progress (current to the single-engine test phase), BOM, costs exceeding ¥8000+), printing status, accessory procurement status.
Address of the project:
However, the cost of both is much higher than the original sales price, and interested classmates can go around.
Why not become a monkey?
To tell you the truth, why a rabbit, not a dog or a human?
In fact, the image of the neck is essentially a natural development according to the body proportions of the robot, the structure of the mouth and the shaking patterns. Since walking would have been a bit of a bumpy step, sex actually becomes a bump.
Moreover, they did not intentionally hide the mechanical structure completely. Although the appearance is lovely, the user can still clearly see that it is a robot.
Interestingly, each monkey has its own voice personality.
At the first launch, the Microduck acquired its own voice identity and has continued to use it ever since.
It's not just a development kit.
Even if you don't write the code at all, you can use it as an electronic pet to play with, because when it comes out of the factory, it's already built in a bunch of well-trained behaviors.
It's all about having fun, folding, and practicing AI.