Clear a Saturday, open a terminal, and you can start teaching a walking robot before the hardware even ships. That is the odd shape of the Microduck launch: Hugging Face and Pollen Robotics say the first units leave the warehouse ahead of Christmas 2026, but the simulation stack the robot learns inside is going on GitHub first. For a thesis group or a school robotics club, the waiting period is the project.
What Hugging Face and Pollen actually built
Microduck stands 25 cm tall and weighs 780 grams, with 15 degrees of freedom to move around. Out of the box it waddles, sits, crouches, picks itself up after a fall, and roller skates. There are no arms. An articulated beak does gripper duty for small objects instead, which keeps both the joint count and the bill of materials down. Sensing is a wide-angle camera, an 8x8 time-of-flight LiDAR array for ranging, and two IMUs for balance and body-pose estimation. Microphones, a speaker, Wi-Fi, Bluetooth and a pair of NFC antennas round it out, so tagged physical objects become fair game. Everything runs on the robot, with no internet connection required.
Parts and cost reality
Preorder pricing is $399, roughly PHP 23,000 at current rates. That is a fraction of what a research-grade biped normally runs, though it is still real money for a student org, so it is worth knowing what the peso buys. Power is a removable 2,600 mAh NP-F550 camera battery, the same cell sitting in half the video rigs in your AV room. Runtime lands around an hour, which means a second NP-F550 is the first accessory to budget for: swap packs and your training session never stops for a recharge.
Spend your Sunday in simulation
Reinforcement learning on physical hardware is a bad idea, because a policy needs thousands of failed attempts and a 780-gram robot will not survive them. The workflow here is sim-to-real: build the environment, train the policy in simulation, then flash the trained behaviour onto the robot and watch where reality disagrees. Pollen says the SDK covers that whole loop, from environment authoring through robot control and deployment. The tricky part is always the reality gap, and the fix is boring iteration: tweak the sim, retrain, redeploy. Nothing stops you from starting that loop now on a laptop. Read the full announcement at Hackster, then watch the Hugging Face GitHub org for the SDK drop so your team is already fluent in the simulator on day one.
Originally published on blog.circuit.rocks.
Top comments (0)