Microduck

Open on YouTube

External video · Pollen Robotics on YouTube

Microduck is a compact, duck-shaped bipedal robot designed to make physical AI more accessible. Standing approximately 25 cm tall and weighing around 780–800 g, it packs 15 motors, onboard sensing, an articulated beak, and an open-source software stack into a desk-friendly platform built for movement, experimentation, and play.

Its appeal extends well beyond its distinctive appearance. Microduck gives developers and robotics enthusiasts a practical way to train AI-controlled behaviors in simulation and deploy them on real hardware. Its appearance in NVIDIA's GTC Berlin keynote in October 2026 further highlighted a broader shift: advanced learning techniques are increasingly being explored on robots small enough to sit on a desk.

Microduck was developed by Pollen Robotics, the French robotics company acquired by Hugging Face in April 2025. Unlike conventional desktop companions that primarily turn their heads, display expressions, or respond to conversation, Microduck is built around physical movement.

Its 15 motors drive an articulated body that includes the legs, head, and neck. With onboard sensors and learned control policies, it can walk, sit, crouch, recover from many common falls, and pick up light objects with its articulated beak. An optional wheel kit enables roller-skating experiments.

The result is a robot that makes physical AI tangible. Instead of merely watching an AI model generate text on a screen, users can observe a learned policy produce real movement in the physical world.

Microduck's capabilities are best understood as five connected areas — locomotion, sensing, simulation-based training, the transfer of learned policies to real hardware, and the accessibility of the whole workflow.

• Bipedal locomotion and fall recovery — balancing, walking, crouching, and getting back up on two legs.
• Sensors for physical interaction — camera, depth sensing, IMUs, and an articulated grasping beak.
• NVIDIA-powered simulation and AI training — Isaac Sim, Isaac Lab, and GPU-accelerated parallel training.
• From simulation to reality — the practical gap between a trained policy and real-world behavior.
• An accessible platform for experimentation — an open-source software stack for developers and educators.

Microduck is designed to balance and move on two legs rather than relying on wheels as its primary mode of locomotion. Its motion system combines 15 motors with inertial sensing and distance measurements to support a range of physical behaviors.

Walking, sitting, crouching, and getting back up after a fall are among its demonstrated capabilities. These make Microduck a useful platform for exploring the core challenges of legged locomotion — balance, actuator response, and ground contact.

As with other learning-based robotic systems, real-world performance depends on the specific policy, software version, and operating conditions.

Microduck integrates several sensing and interaction components in a compact form factor:

• Front-facing camera
• 8 × 8 time-of-flight depth sensor
• Two inertial measurement units (IMUs)
• Articulated beak for grasping light objects
• Microphones and speaker
• Wi-Fi and Bluetooth connectivity

The depth sensor provides a coarse grid of distance measurements, while the IMUs track motion and orientation. Combined with the camera and control software, these support experiments in movement, perception, and environmental interaction.

The beak is more than a visual flourish: it functions as a small manipulator capable of picking up selected lightweight objects.

One of Microduck's most interesting aspects is its connection to NVIDIA's robotics software ecosystem.

NVIDIA Isaac Sim provides a simulated environment for evaluating robot behavior. Isaac Lab supports reinforcement learning and related robot-learning workflows, while NVIDIA GPUs accelerate training across many parallel simulations.

In an October 2026 technical demonstration, NVIDIA described training a Microduck locomotion and perception policy across 4,096 parallel simulated environments on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU. The workflow used Isaac Sim, Isaac Lab, Newton, and RTX Spark for simulation, training, and policy evaluation.NVIDIA developer blog

This approach lets developers explore how a robot might respond to ramps, obstacles, and varying terrain before deploying the resulting policy on physical hardware.

Training a robot in a virtual environment does not automatically guarantee identical behavior in the real world.

Physical motors introduce response delays. Battery voltage affects actuator performance. Floors vary in friction, and real sensors produce measurements that never perfectly match simulated observations.

Microduck offers a concrete example of the simulation-to-reality process. Developers can train a policy in simulation, evaluate its behavior, export a compatible model, and deploy it to the physical robot. Subsequent testing reveals where the simulated model and real hardware diverge.Hugging Face blog

This workflow matters because it connects software experimentation with physical outcomes. A successful policy must ultimately work with real actuators, real sensor data, and real-world contact — not merely look convincing in a virtual environment.

Microduck is intended for more than passive demonstration. Its open-source software stack lets users explore existing behaviors and develop their own.

The stack provides an entry point for developers interested in reinforcement learning, locomotion, perception, and physical AI. The robot can also serve as an educational platform for understanding the relationship between simulated training environments and real robotic systems.

The distinction between open-source software and open-source hardware is important. Pollen Robotics' published specifications make clear that the software stack is open source, but the mechanical and electronic design files are not.

Although Microduck is closely associated with simulation and AI training, the robot that ships to customers is a self-contained physical device. Its onboard compute handles sensing and control locally, while the heavier training work happens elsewhere.

The platform pairs a Rockchip RK3566 processor with an AI accelerator, 1 GB of RAM, and 32 GB of storage. Power comes from a removable NP-F550 battery, giving approximately one hour of runtime depending on use. Connectivity is handled over Wi-Fi and Bluetooth.

That separation — training on external GPUs, execution on modest onboard hardware — is central to how Microduck works, and it is a pattern that appears across a growing number of learning-based robots.

At launch, Microduck was offered for preorder at an introductory price of US$399 before taxes and shipping. Pollen Robotics targeted initial deliveries before Christmas 2026, with availability and schedules dependent on region and order timing.

Buyers should check the official store for current pricing, shipping estimates, and regional availability.Official product page

Microduck is not intended to compete with full-size industrial robots or general-purpose humanoid platforms. Its value lies in making selected robotics challenges easier to explore at a smaller, more accessible scale.

For developers, it provides a physical platform for investigating learned locomotion and simulation-to-reality transfer. For educators and enthusiasts, its compact form makes abstract concepts — reinforcement learning, balance control, sensing, and light manipulation — easier to observe and experiment with.

Its connection to NVIDIA's simulation and training ecosystem also illustrates an important point: progress in robotics does not depend solely on placing a powerful processor inside every machine. Much of the intelligence behind a robot's behavior can be developed, tested, and refined using external computing resources before the resulting policy is deployed to the robot's onboard hardware.

That separation between training infrastructure and physical execution is an increasingly central part of modern robotics development.

Microduck's most interesting feature may not be its duck-like appearance or its ability to perform a somersault. It is the relationship between a small physical machine and the much larger computational process used to teach it new behaviors.

The robot illustrates how simulation, reinforcement learning, hardware design, and developer tools can work together to bring AI out of the purely digital environment and into the physical world.

It is not a fully autonomous general-purpose robot, and its small size limits what it can carry and manipulate. Nevertheless, it offers a concrete example of how physical AI development is becoming more accessible to individuals and smaller teams.

For Robodoor, Microduck is worth watching not simply as a novelty robot, but as a compact demonstration of the tools and workflows shaping the next generation of consumer and developer robotics.

As with any platform at this scale, Microduck's strengths and its constraints come from the same design decisions.

The information on this page is based on Pollen Robotics' published product information and NVIDIA's technical documentation.

• Pollen Robotics — Microduck product page
• Pollen Robotics — Microduck press kit
• NVIDIA — Using AI Agents to Teach Robots Like Microduck to Walk in Simulation
• Hugging Face — Microduck Isaac Lab sim-to-real