Superpower Daily
Hugging Face Sells a $399 Microduck Robot With Its Training Sta...Google DeepMind Puts Gemini Flash Lite in a Double-Blind Test t...Databricks Says 300M Chart-JSON Pipeline Tops Four Multimodal B...Socure Raises $156M at $5.2B, Buying Fravity to Automate Fraud...Harness Launches Agent-Ready Repository and AI Review for High-...Businesses Turn ‘No AI’ Signs Into Ads, With Some Drawing Tens...Hugging Face Sells a $399 Microduck Robot With Its Training Sta...Google DeepMind Puts Gemini Flash Lite in a Double-Blind Test t...Databricks Says 300M Chart-JSON Pipeline Tops Four Multimodal B...Socure Raises $156M at $5.2B, Buying Fravity to Automate Fraud...Harness Launches Agent-Ready Repository and AI Review for High-...Businesses Turn ‘No AI’ Signs Into Ads, With Some Drawing Tens...
Productspublished

Hugging Face Sells a $399 Microduck Robot With Its Training Stack on GitHub

The launch pairs inexpensive hardware with a stated simulation-to-robot workflow, while leaving data handling to the software users install.

By 2 min read
Hugging Face Sells a $399 Microduck Robot With Its Training Stack on GitHub
Sign in to save and react

Listen to this story

The audio brief

About 1:27
0:001:27
Read transcript
Hugging Face is preparing a $399 duck-shaped robot that developers can teach, retrain, and redeploy—rather than a gadget locked to one purpose. Microduck is 25 centimeters tall, scheduled to ship before Christmas, and built around a public reinforcement-learning toolkit. Its SDK, simulator, and full training stack are available on GitHub. The intended workflow starts in simulation: developers train a behavior, deploy it to the physical robot, fine-tune it after that first run, then retrain and redeploy. That creates a repeatable loop between software experiments and a real machine. The robot’s listed abilities include waddling, crouching, recovering after a fall, and roller skating. It can also pick up objects weighing up to 800 grams with its beak. A camera, lidar, and two inertial measurement units, or IMUs, help it perceive movement and surroundings. The launch extends Hugging Face’s physical-AI push after its April 2025 acquisition of Pollen Robotics, which was followed by the Reachy Mini products. The important caveat is privacy. Open-source code may offer more auditability, but it is not a data-handling guarantee: installed applications can access cameras and microphones and may transmit captured data to outside services. So the immediate question is not whether Microduck can learn new tricks, but how developers handle the sensor data once those tricks move from simulation into homes, labs, and products.

Story brief

3 key points

Hugging Face and Pollen Robotics are preparing a $399, 25-centimeter Microduck robot that lets developers train reinforcement-learning behaviors in simulation, deploy them to hardware, then fine-tune and repeat. The SDK, simulator, and training stack are public on GitHub, positioning the device as an affordable physical-AI development platform rather than a fixed-purpose gadget. Microduck is slated to ship before...

  1. 01

    Microduck can pick up objects weighing up to 800 grams and perform locomotion behaviors including waddling, crouching, recovery, and roller skating.

  2. 02

    Its sensor suite includes a camera, lidar, and two inertial measurement units; applications may transmit captured data to external services.

  3. 03

    The launch extends Hugging Face’s physical-AI strategy after acquiring Pollen Robotics in April 2025 and launching Reachy Mini products.

Hugging Face’s Microduck pairs a $399, 25-centimeter duck-shaped robot with a public reinforcement-learning toolkit. The promise is a route from simulation to deployment on a physical machine. The launch supports that workflow and a set of physical behaviors, but privacy still depends on the applications users install on its sensor-equipped hardware.

Microduck is scheduled to ship before Christmas. Hugging Face CEO Clem Delangue describes it as an open-source robot that developers can teach new tricks with reinforcement learning. Its SDK, simulation environment and full training stack are available on GitHub.

The product is a training loop, not just a robot

Pollen Robotics says developers can train behaviors in simulation and deploy them directly to Microduck. After that first deployment, they can fine-tune the robot, retrain it and redeploy it. The stated proposition is a repeatable development cycle that connects the simulator and the physical device.

A camera, lidar and motion sensors underpin the tricks

Microduck perceives its surroundings with a camera, lidar sensors and two inertial measurement units, or IMUs, which measure movement. Its listed behaviors span locomotion and simple manipulation rather than one fixed task.

Listed abilities

  • Movement: waddle, crouch, recover after falling and roller skate.
  • Handling: pick up objects with its beak weighing up to 800 grams.

Open source is not a privacy setting

Delangue has argued that robots running open-source models give users more auditability and control than black-box systems. But open source is not a data-handling guarantee: installed applications can access robot cameras and microphones and may send that data to external services, depending on how they are built. Application design, not the availability of code alone, sets that privacy boundary.

The next hardware product after Pollen Robotics

Microduck follows Hugging Face’s April 2025 acquisition of French startup Pollen Robotics, which was intended to build affordable, open-source AI hardware. The companies later launched Reachy Mini. They now sell a $499 Raspberry Pi-powered Reachy Mini and a $399 Reachy Mini Lite that runs from a Mac or PC.

Sources

  1. techcrunch.comHugging Face is selling a cute $399 open-source duck robot, Microduck | TechCrunch

Related coverage

Continue reading