Hugging Face Just Shipped a $399 Robot Duck. Here’s What It Actually Means for NVIDIA.

Microduck is not an NVIDIA robot. Per Automate's teardown, it runs on a Rockchip RK3566, and it launched before NVIDIA's $12.9 billion Hugging Face deal was official. But read the training docs and you'll see why NVIDIA is the company that benefits most.

The Problem It's Solving

Teaching a robot to move is expensive and messy. The robot has to try things, fail, and try again. On a full-size humanoid, Pollen Robotics says, every bad attempt can be costly, hard to reset, or unsafe to run outside a lab.

That leaves most developers training legged behaviors purely in simulation, with no cheap way to test whether the policy survives real floors and real motors. Sim-to-real is the hard part of physical AI, and the hardware to practice it has been priced for labs.

Microduck is Pollen's answer: a robot small and cheap enough that a failed behavior ends with a duck on the floor, not an incident report.

How Microduck Actually Works

Microduck is a 25 cm biped under 800 g. It has 15 motors, a camera, a small depth sensor, two IMUs, and an articulated beak that works as a gripper. It walks, sits, crouches, recovers from many common falls, and roller-skates. It can't talk. It makes creature sounds, and each unit gets its own audio identity on first wake.

The robot ships with learned behaviors you can trigger with a gamepad, so it's useful before you write any code. Underneath, the whole stack is open. The main repository holds the SDK and robot software. Microduck RL holds the reinforcement learning and sim-to-real tools. Pollen's pitch is that Reachy Mini is a platform for AI that interacts, while Microduck is a platform for AI that acts.

The workflow is the standard sim-to-real loop. You train a policy across many simulated ducks, export it, deploy it to the real robot, watch what breaks, and go back to simulation. There is also a browser simulator for running pretrained policies before you own the hardware.

Here is the NVIDIA part. According to a Chinese-language breakdown of the official training docs, local training needs a CUDA GPU because the pipeline uses MuJoCo Warp. Without one, you can submit the same job to cloud GPUs through Hugging Face Jobs. A community developer ported the training environment to AMD GPUs two days after launch, but that port is unofficial, only partly working, and not yet verified on a real robot. The default path runs on CUDA.

What Developers Are Actually Using It For

The use cases so far are narrow and technical, which is a good sign. Seeed Studio published a Jetson walkthrough that covers PPO training in MuJoCo on a Jetson Orin NX 16GB, visualizing checkpoints, running the official ONNX policy with keyboard control, and building custom motions.

Pollen's own team is pushing the robot's perception. Two October posts cover SLAM with the 8×8 depth sensor and contact odometry from two feet and an IMU, both running fully onboard. Pollen also says Microduck is about ten times more fun in groups, which makes multi-robot experiments (races, football) practical without a room full of expensive hardware.

Why This Is a Bigger Deal Than It Looks

The sequencing matters. Fortune's report on the NVIDIA talks is dated August 27, the same day Pollen published the Microduck announcement. NVIDIA then confirmed the deal on September 3. Some coverage says Hugging Face turned in Microduck as its first assignment after the acquisition. That's wrong. The robot predates the announcement.

Still, the strategic logic holds. Jensen Huang wrote that Hugging Face will stay an open platform and that NVIDIA compute will not be required to build on or deploy through it. Microduck's default training path doesn't contradict that, since the code is open. But a $399 robot whose official pipeline assumes CUDA is a very cheap way to recruit the next generation of robotics developers into NVIDIA's stack. Automate makes the same point: the real opportunity is developer distribution.

Robotics is where NVIDIA wants its next platform to be, and Hugging Face brings the community that shares policies the way it already shares models. Pollen frames it the same way: physical behaviors should be as easy to share as models, along with the environments and training recipes to reproduce them.

Availability and Access

Microduck is open for pre-order at an introductory $399 before taxes and shipping, in Cream, Graphite, Lavender, and Sky. First deliveries are targeted before Christmas 2026 in North America, Europe, and the UK. The software is open under Apache 2.0 and available today. As for the acquisition, reports describe it as expected to close in the first half of 2027, subject to regulatory review.

Microduck won't solve robotics, and Pollen says as much. But the company that owns the GPUs, the model hub, and now the cheapest on-ramp to physical AI is assembling a full stack, one duck at a time.

Follow for more coverage on MCP, agentic AI, and AI infrastructure.