Product2 distinct publishers3 min readPublished
At $399 the Microduck is small enough and cheap enough to break on purpose, which turns embodied-AI experiments into an expense-report purchase rather than a capital project, once you know when it ships.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The line item worth arguing about is the fall. Engadget makes the point plainly: a large humanoid taking repeated tumbles carries both safety and financial cost, while this machine picks itself back up and does not hit the floor with much impact [7]. Fifteen motors, a camera, a small LiDAR sensor and two IMUs in a 25cm body under 800g [3] means the failure mode you most want to study, a balance policy losing its footing and recovering, is something you can run until it stops being interesting. Hugging Face says the motors let it walk, sit, crouch, get back up from many common falls and roller-skate [5], and the reinforcement-learning tools sit in their own repo alongside the SDK [8].
Hugging Face's stated case for owning several is that it is "about ten times more fun", and that the team got some playing soccer [13]. The operational case is wall clock. Each unit runs about an hour of continuous use on its 2,600mAh battery [10], so a team buying units is really buying hours: five of them cost $1,995 [17] and yield roughly five robot-hours per charge cycle [18]. That is a number most teams can put on a purchase order without opening a capital request, which is the part of this announcement that changes anyone's planning.
Separate the pitch from the work. The pitch is character: no words, only sounds, an audio identity generated the first time the robot wakes and kept for life [14], and four colorways so that a group looks like a collection of different characters rather than identical machines [16]. The work is duller and better evidenced by the demo footage the Verge describes, in which the robot picks up socks and markers, kicks a ball around and rolls on tiny skates [15]. Grasping a sock with an articulated beak is a real manipulation result at a real scale, and it is the thing a buyer should be judging.
Two axes decide whether this belongs on your bench. Axis one is whether your open question needs contact with an uncooperative floor or a simulator will do. Axis two is whether the answer has to hold at your target hardware's mass. If contact matters and mass does not, $399 for a machine explicitly aimed at developers experimenting with obstacles and accidents [6] is the cheapest instrument on the market for that question. If mass matters, nothing here transfers, because a sub-800g body with a beak teaches you about recovery and small-object grasp, and no price fixes the gap to a load-bearing platform.
The measure to hold yourself to afterwards is not hours of play. It is time from unboxing to a behavior you retrained yourself rather than one of the pre-learned behaviors that ship in the box [8], and whether anyone on the team asks for a second unit. Engadget notes the announcement post is light on technical detail [19], so budget the first week for reading code rather than running trials.
Ranked by verification strength, evidence, and original report placement.
Pollen Robotics, which Hugging Face acquired last year, opened pre-orders for the Microduck robot at $399, with four colorways offered: Cream, Graphite, Lavender and Sky.
Microduck stands 25cm tall, weighs under 800g, and is a biped with 15 motors, a built-in camera, a small LiDAR sensor, two IMUs and an articulated beak able to pick up objects.
The Verge describes Microduck as a one-eyed biped standing just under 10 inches tall, running a Rockchip RK3566 processor, with a camera, motion sensors, LiDAR and articulated legs, head and neck.
According to its maker, motors spread across the legs, head and neck allow Microduck to "walk, sit, crouch, get back up from many common falls, and even roller-skate".
The main Microduck repository is available today and includes the SDK and robot software, with a Microduck RL repo containing reinforcement learning tools; the robot can be driven with a gamepad or made to follow a laser dot out of the box, and pre-learned behaviors let it react to surroundings without coding. The software stack is open source.
Microduck does not use words and communicates through sound via a built-in microphone and speaker; each robot gets its own audio identity the first time it wakes up, tied to that individual robot and kept for life, and owners are meant to think of it as a creature rather than an assistant.
Distinct publishers with included, body-backed reporting in this cluster.
2 articles · August 27, 2026
1 article · August 27, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
XPeng's $900m robot spin-out comes with a seven-year clock and a $1bn put2 distinct publishers
invest
OpenAI's own timeline: twelve days from agent attack to knowing it was them1 distinct publisher
product
Ox Alpha was GLM-5.3-Flash, and the number that decides displacement is 18 billion1 distinct publisher
build
Shanghai AI Lab's 397B science agent shipped in July; the paper explaining it landed August 131 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Vendor-sourced specs, no hands-on
Specs and capabilities are consistent across two publishers and specific enough to act on (25cm, 15 motors, LiDAR, two IMUs, RK3566, 1GB RAM, 32GB storage, 2,600mAh/~1h). But both accounts trace to one Pollen Robotics blog post and vendor video; neither publisher tested the robot, Engadget notes the announcement is thin on technical detail, and the shipping year is inconsistently reported. Repo availability is asserted rather than demonstrated in the sources.
Pre-orders open, nothing shipped
Adoption is limited to a commercial launch event: pre-orders opened and repositories were published on announcement day. No units have shipped, no pre-order counts, waitlist size, GitHub traction, benchmark results or third-party deployments appear in any source. The only user-scale signal is the vendor's own internal remark that several robots at once are 'about ten times more fun'.
Charming demo ahead of verified capability
Positive but moderate. The vendor's capability list (walk, sit, crouch, recover from many common falls, roller-skate) and the creature-with-a-lifelong-voice narrative are doing more work than any verified result, the announcement is admittedly thin on technical detail, and money is being taken now for hardware that may not arrive until late 2026. The gap is not larger because the falsifiable parts — price, colorways, dimensions, sensors, battery hour, public repos — are concrete, cheap to check on delivery, and consistently reported.
Pre-order marketing relayed from vendor blog
The originating document is a vendor announcement whose purpose is to convert readers into pre-order deposits, and both publishers build their pieces from it, quoting the maker directly on capability, fun and personality. Gadget outlets also have an audience incentive toward a charming duck robot. No paid placement, affiliate arrangement or reviewer-unit relationship is disclosed in the sources, so this reflects visible structural incentive rather than any established conflict.
Solid on the announcement, thin on outcomes
High confidence that the pre-order, pricing, colorways, spec sheet and open-source repos are as reported, since two publishers agree and the details are precise. Low confidence on anything downstream: capability under real use, delivery date and year, licence terms, and adoption. Effective source count is two publishers on one vendor post, one of them duplicated in the cluster.