Que s'est-il passé
36Kr reports that Hugging Face and Pollen Robotics released Microduck, a small two-legged robot priced at $399. The robot can walk, kick a ball, skate with optional rollers, use its opening beak as a gripper, pick up small objects and recover after falling. Its software package includes pre-trained motion policies, an SDK, a simulation environment and reinforcement-learning tools.
36Kr reports that Microduck is 25 centimeters tall, weighs less than 800 grams and costs $399, or about 2,800 Chinese yuan. The small robot has two legs and a movable beak that functions as a gripper. According to the report, demonstrations show it walking, sitting, standing, squatting, kicking a small ball and skating when rollers are attached to its feet. It can also lower its body to pick up objects such as markers and socks, carry them away and recover independently after falling. These capabilities are presented as demonstrations, not as independently verified tests.
The central product feature, according to 36Kr, is not the duck’s choreography but its openness. Microduck ships with multiple pre-trained motion policies, while Pollen Robotics is reported to be releasing an SDK, simulation environment and reinforcement-learning tools. Developers can create virtual copies in the MuJoCo physics simulator, train them through repeated attempts to balance and walk, then export a neural-network policy to the physical robot. The report describes this as a simulation-to-reality workflow.
36Kr reports that the official training example can run 4,096 simulation environments simultaneously and that the company estimates a usable walking gait can be trained in roughly one to two hours on a CUDA-capable Nvidia GPU. Developers without suitable local hardware can reportedly submit training jobs to cloud GPUs through Hugging Face Jobs. The source cautions that successful simulation does not guarantee successful real-world behavior because floor friction, motor condition and the robot’s center of gravity differ from the simulated environment.
The report places Microduck within Hugging Face’s broader robotics effort. 36Kr says Hugging Face began the LeRobot open-source robotics project in 2024, later released the SO-100 and SO-101 robotic arms, and acquired Pollen Robotics in 2025. It reports that the company subsequently introduced the $399 Reachy Mini and is using lower-cost physical systems to give developers hardware on which they can repeatedly experiment. The source also says that, around the Microduck release, a community effort to adapt training for AMD GPUs was incomplete and had not been verified on a real robot.
Détails de la source: eu.36kr.com ↗
Pourquoi c'est important
Microduck lowers the cost of experimenting with embodied AI and connects physical hardware to Hugging Face’s , and training ecosystem. 36Kr reports that developers can train virtual copies in simulation, use local CUDA hardware or cloud GPUs, and deploy learned policies to the physical robot, although the source contains no independent performance testing.
The reported price changes who can participate in embodied-AI experimentation. 36Kr contrasts Microduck’s $399 price with research robots that can cost tens of thousands of dollars or more. A cheaper machine can make physical trial and error more accessible to individual developers, students, laboratories and hobbyists, while reducing the financial consequences of falls and failed experiments. That is a practical difference from software-only experimentation, where a model can be downloaded without purchasing a physical platform.
The product also illustrates Hugging Face’s attempt to extend its to robotics. 36Kr reports that LeRobot is intended to bring models, datasets, training methods and hardware closer together, addressing a field in which laboratories often use separate technology stacks. If the tools are genuinely reusable, developers could adapt behaviors rather than treating a robot as a fixed consumer device. The source does not establish how broad the current community, documentation or hardware compatibility is beyond the examples it describes.
Data and collaboration are another potential benefit. Citing an IEEE Spectrum report, 36Kr says the number of robot datasets on LeRobot rose from 1,145 at the end of 2024 to more than 58,000, making robotics the largest category on Hugging Face Hub. 36Kr also reports that Nvidia, Alibaba, research institutions and home enthusiasts have uploaded models. Those figures and descriptions are not independently confirmed in the source, but they indicate the scale of the ecosystem Hugging Face hopes to connect to inexpensive hardware.
The project also exposes a tension between open software and platform dependence. 36Kr reports that the default Microduck training route relies on CUDA through MuJoCo Warp, while an AMD-compatible community migration was only partly complete and had not been tested on the real robot. This means the device may be inexpensive while the most straightforward development path still depends on particular computing hardware. The source says AMD-optimized LeRobot models exist, but it provides no evidence that the full Microduck workflow is hardware-neutral.
Mécanisme interactif : comment cela fonctionne réellement
Explorez de manière interactive la technologie sous-jacente à ce développement.
crm_get_transaction(id='4092').An agent must create a draft calendar event for Tuesday at 2 p.m. Which evidence would establish the requested result?
Que regarder ensuite
The key questions are whether Microduck can be purchased and used at the reported price, how reliably simulated behaviors transfer to real hardware, and whether the community can support alternatives to the CUDA-based training path. 36Kr also says a reported Nvidia acquisition of Hugging Face remains officially unconfirmed.
The first issue is the gap between a documented demonstration and a dependable development platform. 36Kr reports specific movements and an estimated training time, but the source includes no independent tests of walking stability, object handling, skating, recovery, battery life, durability, noise, safety or repeatability. Developers will need evidence about how often policies trained in simulation work on real units and how much adjustment is required after deployment.
Availability and real-world cost also remain important unknowns. The report gives a $399 price, but it does not establish shipping timing, regional availability, production volume, accessories, taxes, support terms or whether the listed price includes everything needed for training. 36Kr says the robot became highly popular and attributes a pace of one unit sold every four seconds, but it supplies no sales period, underlying data or independent confirmation for that claim.
The software ecosystem will determine whether Microduck becomes more than a low-cost novelty. Watch for completed AMD or other non-CUDA training paths, reproducible community policies, clearer documentation and evidence that users can modify the hardware without losing reliability.
36Kr also discusses reports of a possible $12.9 billion Nvidia acquisition of Hugging Face, but says neither company has officially announced the transaction. Its status, if it changes, could affect how the open-source robotics strategy develops; it should not be treated as a completed deal based on this source.