Why
The reason this item is worth a slot is the price point, not the duck. Reinforcement learning on physical hardware has historically needed a research lab: a robot, a simulator matched to that robot's dynamics, and a pipeline to move a trained policy from sim to real without the "reality gap" destroying it. Pollen ships all three pieces bundled and open: the SDK, a simulation environment, and a full RL training stack, all on GitHub. That turns "physical AI," a phrase Hugging Face CEO Clem Delangue used to describe Microduck as "an open-source robot you can teach new tricks with reinforcement learning," from a research-lab exercise into something a developer can run against a $399 piece of hardware. The target audience the coverage names is developers interested in physical AI and world models — the same audience Jayverse's Theory and Eng sections are written for.
How it works
Body and sensors
15 motors distributed across the legs and beak give Microduck its degrees of freedom; the sensor stack (front camera, 8x8 LiDAR, two IMUs, mics, speaker) is enough for basic perception and balance but modest by robotics-lab standards — this is a hobbyist and developer platform, not a research-grade humanoid. NFC lets it recognize tagged accessories and objects, and it connects over Wi-Fi and Bluetooth, with a game controller as the default input device.
What ships trained, out of the box
Seven trained moves come pre-loaded and are driven by the bundled game controller: walk, sit/stand, kick, grab, roller-skate and self-recovery from a fall. Coverage frames the falling-and-getting-up behavior as a deliberate demo of the RL loop rather than a failure mode — the robot is expected to fall and shows the recovery policy working. Multiple Microducks can be gathered for races and football, and the articulated beak is the robot's manipulator for picking up NFC-tagged objects (within that roughly 100 g lift limit).
The open part is the software stack, not confirmed to be the hardware
"Software fully open source" is the exact claim in Pollen's own description: the SDK, the simulator, and the RL training code are public on github.com/pollen-robotics/microduck. What is not established from the material gathered here is whether the mechanical design (CAD, bill of materials, PCB schematics) is open the way the software is — treat hardware openness as unconfirmed rather than assume it follows from "open source."
Where it fits next to Reachy Mini
Microduck is the third product in Pollen's post-acquisition lineup: Reachy Mini ($499, Raspberry Pi powered, standalone compute) and Reachy Mini Lite ($399, needs a host Mac or PC) came first. Microduck is priced at the Lite tier but is a legged, mobile platform rather than the Reachy Minis' stationary desk-companion form factor — a meaningfully different sim-to-real problem (locomotion and balance versus arm/head manipulation from a fixed base).
Business context, unconfirmed
Press coverage of the Microduck launch also noted that Hugging Face is reportedly in talks to be acquired by Nvidia at a $13B valuation. That report is unconfirmed and orthogonal to Microduck's specs; it is included here only because it appeared in the same coverage window and explains some of the attention the launch got.
Where it lands in Jayverse
- Theory: a physical instance of the sim-to-real gap. Microduck's GitHub stack (simulator + RL trainer + real-robot deployment, then retrain-and-redeploy) is a concrete worked example for any Theory note on policy transfer, domain randomization, or why a policy trained in sim needs a recovery behavior for the real world's noise.
- Game: a reference for a physical mirror of the Unity street. The seven-trained-move, game-controller-driven interaction model (bounded behavior set, controller as default input, autonomous mode as an upgrade path) is a useful shape to borrow for how NPCs or player-controlled characters in the Unity street could expose a small fixed move set plus an autonomous fallback.
- Dark Horse: robotics stays a watch item, not a build. This is the kind of low-cost, open-source robotics signal alice-tech has flagged before; Microduck itself is not a Jayverse build candidate, but it is evidence the hobby-robotics-plus-RL category is maturing fast enough to keep watching.
- Eng: a ready interview answer for "what is physical AI." Microduck is a concrete, citable example — bundled sim-to-real RL stack, real hardware, $399 — for explaining "physical AI" in a dev/team-lead interview without hand-waving.
Verified and unverified
Verified on 2026-09-19: Pollen Robotics is a real French robotics company acquired by Hugging Face in April 2025; Reachy Mini ($499) and Reachy Mini Lite ($399) are real, documented Pollen/Hugging Face products; a GitHub repository at pollen-robotics/microduck and a product page at pollen-robotics.com/microduck are named consistently across the source material; the general shape of the claim (open SDK, simulator and RL training stack for a small legged robot) is consistent with Hugging Face's existing open-source robotics posture. Taken from the video description and the TechCrunch/Register/MarkTechPost coverage and not independently checked here: the exact 25 cm / 780 g dimensions, the 15-degrees-of-freedom and 8x8 LiDAR specs, the ~100 g lift-capacity figure attributed to Santiago Pavon, the seven-trained-move list, the 1.7 million view count, the $399 pre-order price and Christmas-2026 delivery target, and the reported Nvidia acquisition talks at a $13B valuation (explicitly unconfirmed by the source coverage itself). Whether Microduck's hardware design (not just its software) is open source is not established by any source gathered and should be treated as unknown.
Sources: Pollen Robotics YouTube, "We made a new robot," Aug 27 2026 · pollen-robotics.com/microduck · TechCrunch, "Hugging Face is selling a cute $399 open-source duck robot: Microduck" · The Register, "Hugging Face offers $399 robot duck to help you quack the AI code".
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| sim-to-real | 시뮬레이션에서 실제로(학습된 정책을 시뮬레이터에서 실기로 옮기는 과정) · 강화학습 로보틱스의 핵심 용어. "the sim-to-real reinforcement-learning loop" |
| reality gap | 리얼리티 갭(시뮬레이션과 실제 물리의 차이로 정책이 무너지는 현상) · sim-to-real 전이의 실패 원인을 가리킬 때. "without the 'reality gap' destroying it" |
| physical AI | 물리적 AI(로봇 등 실제 하드웨어에서 동작하는 AI) · Hugging Face CEO가 Microduck을 설명하며 쓴 표현. "developers interested in physical AI and world models" |
| degrees of freedom | 자유도(관절·모터가 독립적으로 움직일 수 있는 축의 수) · 로봇의 움직임 능력을 세는 표준 단위. "15 degrees of freedom (15 motors)" |
| IMU | Inertial Measurement Unit(관성 측정 장치, 가속도·자세를 재는 센서) · 균형 제어에 쓰이는 센서. "two IMUs" |
| NFC | Near Field Communication(근거리 무선 통신) · 태그된 액세서리·물체 인식에 쓰임. "NFC lets it recognize tagged accessories" |
| LiDAR | Light Detection and Ranging(레이저로 거리를 재는 센서) · Microduck의 8x8 배열 거리 센서. "an 8x8 LiDAR" |
| RL | Reinforcement Learning(강화학습, 보상으로 정책을 학습시키는 기법) · Microduck의 학습·재학습 스택의 핵심. "a full RL training stack" |
| SDK | Software Development Kit(소프트웨어 개발 키트) · GitHub에 공개된 오픈소스 구성 요소 중 하나. "the SDK, a simulation environment" |
| CAD | Computer-Aided Design(컴퓨터 지원 설계, 기계·부품 도면 소프트웨어) · 하드웨어 개방성 미확인을 설명할 때. "CAD, bill of materials, PCB schematics" |
| PCB | Printed Circuit Board(인쇄 회로 기판) · 하드웨어 오픈소스 여부를 논할 때 쓰이는 부품 예시. "PCB schematics" |
| bill of materials | 부품 명세서(제품을 구성하는 모든 부품의 목록) · 하드웨어를 완전히 재현하는 데 필요한 문서. "CAD, bill of materials, PCB schematics" |
| domain randomization | 도메인 랜덤화(시뮬레이션의 물리·시각 조건을 무작위로 바꿔 정책을 더 강건하게 만드는 기법) · sim-to-real 전이를 다루는 Theory 노트의 용어. "domain randomization" |
| policy transfer | 정책 전이(학습된 제어 정책을 한 환경에서 다른 환경으로 옮기는 것) · 시뮬레이터에서 실기로 옮기는 것을 가리킬 때. "a Theory note on policy transfer" |
| fixed base | 고정 베이스(로봇이 바닥에 고정되어 이동하지 않는 형태) · Reachy Mini와 Microduck의 폼팩터 차이를 설명할 때. "manipulation from a fixed base" |
| hobby-budget | 취미 예산 수준의(비싼 연구 장비가 아니라 개인이 감당할 가격대) · 이 항목의 핵심 앵글을 요약하는 표현. "a hobby-budget skill" |
| bundled | 함께 묶여 제공되는(제품에 기본 포함된) · SDK·시뮬레이터·트레이너가 한 세트로 오는 것을 설명. "Pollen ships all three pieces bundled and open" |
| pre-order | 사전주문(정식 출시 전에 미리 주문하는 것) · Microduck의 현재 판매 방식. "Pre-orders are open at $399" |
| orthogonal | 무관한, 별개의(수학의 '직교'에서 온 비유, 서로 영향을 주지 않는다는 뜻) · Nvidia 인수설이 사양과 상관없다고 밝힐 때. "orthogonal to Microduck's specs" |
| hand-waving | 손짓만으로 얼버무리기(근거 없이 대충 설명하는 것) · 인터뷰에서 막연한 설명을 피한다는 의미로. "without hand-waving" |
| watch item | (당장 만들지 않고) 지켜보는 항목 · Dark Horse 섹션에서 빌드 후보가 아니라 관찰 대상임을 표시할 때. "robotics stays a watch item, not a build" |