Pollen Robotics unveiled Microduck on August 27, 2026, with a specification that immediately caught the robotics industry’s attention: a 25-centimeter-tall biped robot with 15 degrees of freedom, onboard reinforcement-learning control, vision, depth sensing, dual IMUs, wireless connectivity, and a launch price of just $399 before taxes and shipping.

The price raises an obvious question.

How can a robot built with 15 DYNAMIXEL-class actuators sell for less than the retail cost of the actuators alone?

At current U.S. retail pricing, a ROBOTIS DYNAMIXEL XL330-M288-T sells for $27.49. Fifteen units would cost $412.35 before adding a processor, camera, depth sensor, battery, PCB, mechanical structure, or any other component.

That comparison is real.

The frequently repeated conclusion that Pollen Robotics must therefore be losing money on every Microduck is not supported by public evidence.

Pollen Robotics has not disclosed its production BOM, supplier contracts, actuator purchasing price, manufacturing cost, tooling amortization, logistics expense, warranty reserve, or hardware gross margin.

What the available evidence does show is more interesting: Microduck appears to be priced far below what an individual developer would spend to reproduce an equivalent robot using components purchased through normal retail channels.

Understanding how Pollen achieved that requires looking beyond retail component prices and examining the architecture, manufacturing economics, and strategic role of Microduck inside the emerging Physical AI ecosystem.

What Is Microduck?

Microduck is a compact biped robot developed by Pollen Robotics, the French robotics company acquired by Hugging Face.

According to Pollen Robotics, Microduck has:

• 15 degrees of freedom

• a height of approximately 25 cm

• a weight below 800 g

• a front-facing camera

• depth sensing

• two IMUs

• onboard computing

• a 50 Hz onboard policy loop

• microphone and speaker

• NFC

• Wi-Fi and Bluetooth

• removable battery power

Its introductory price is $399 before tax and shipping.

Microduck is particularly interesting because it is designed around reinforcement learning and sim-to-real deployment rather than conventional pre-programmed robotic motion.

A developer can train behaviors in simulation, deploy policies to the physical robot, observe the results, modify the model, and repeat the process.

That makes Microduck less comparable to a traditional robot toy and more comparable to a compact Physical AI development platform.

The Microduck BOM: What We Know

There is currently no official public production BOM for Microduck.

Pollen Robotics has not released a manufacturing-ready package containing the full electrical schematic, PCB design, mechanical CAD, Gerber files, production tolerances, supplier list, and component part numbers.

Its official press documentation explicitly states that the open-source claim applies to the software stack. The mechanical and electronic design files are not released as open-source hardware.

However, Pollen Robotics has published enough engineering documentation, software, simulation assets, and runtime information to reconstruct much of the hardware architecture with reasonable confidence.

Current Microduck Hardware Map

Component	Known or Most Likely Specification	Qty.	Status
Actuators	ROBOTIS DYNAMIXEL XL330 family	15	Strongly supported
Exact actuator variant	XL330-M288-T	15	Highly likely
Main processor	Rockchip RK3566	1	Confirmed
RAM	1 GB	1	Confirmed
Storage	32 GB	1	Confirmed
Compute platform	Radxa Zero 3-class architecture	1	Strongly supported
Depth sensor	8×8 multi-zone Time-of-Flight	1	Confirmed
ToF sensor	STMicroelectronics VL53L8CX	1	Strongly supported
Camera	Front-facing camera	1	Confirmed
IMUs	Head IMU + body IMU	2	Confirmed
Audio	Microphone + speaker	1 set	Confirmed
NFC	Two NFC antennas	2	Confirmed
Wireless	Wi-Fi + Bluetooth	Integrated	Confirmed
Battery	NP-F550-format, 2600 mAh	1	Confirmed
Mechanical system	Molded polymer, brackets, fasteners and joints	Full set	Confirmed
Accessories	Controller + USB-C cable	1 set	Confirmed

The remaining uncertainty lies mostly in the exact production part numbers and supplier pricing.

The Biggest Cost Driver: 15 DYNAMIXEL Actuators

The actuator system dominates the economics of Microduck.

Pollen Robotics officially specifies 15 degrees of freedom, and its robot-control architecture describes 15 servos operating over the robot’s control bus. The runtime reads the devices and writes 15 joint targets during its 50 Hz control loop.

Official engineering material also points to the DYNAMIXEL XL330 family.

The exact XL330 variant has not been prominently disclosed in Pollen’s consumer-facing specification, although technical reconstruction strongly points toward the XL330-M288-T.

That distinction matters when presenting a production BOM. The XL330 family is well supported by the available evidence; the complete M288-T suffix should still be treated with slightly more caution.

Retail Economics

ROBOTIS currently lists the XL330-M288-T at:

$27.49 per unit in the United States.

For 15 actuators:

15 × $27.49 = $412.35

Microduck:

$399

This means an American consumer buying 15 comparable actuators individually would spend more on the actuators than on an entire Microduck.

And the robot still needs:

• compute

• RAM

• storage

• camera

• depth sensor

• dual IMUs

• audio

• wireless connectivity

• battery

• power management

• PCBs

• cables

• structural parts

• injection molding

• assembly

• testing

• packaging

• accessories

This is the single most important number in understanding Microduck’s pricing.

It is also where simplistic BOM analysis often goes wrong.

Retail price is not OEM cost.

A company purchasing thousands or tens of thousands of components under a commercial supply agreement can operate under a fundamentally different cost structure from an individual buying a servo from an online store.

Pollen’s actual DYNAMIXEL procurement price is not public.

Without that number, no outside analyst can calculate Microduck’s exact hardware margin.

The Compute Platform: Rockchip RK3566 and Radxa

Pollen Robotics officially identifies the Microduck compute platform as:

Rockchip RK3566

with:

1 GB RAM and 32 GB storage.

The RK3566 is a quad-core Arm Cortex-A55 system-on-chip that has become common in compact Linux devices, edge computing systems, and embedded hardware.

The more interesting question is whether Microduck literally contains a commercial Radxa Zero 3W board.

Official Pollen engineering records show Microduck software being built, deployed, and tested on real Radxa Zero 3W hardware. One bring-up document explicitly states that the observed results came from a real Radxa Zero 3W rather than simulation.

Radxa’s own specification is an extremely close fit:

• Rockchip RK3566

• quad-core Cortex-A55

• LPDDR4

• onboard eMMC

• Wi-Fi 6

• Bluetooth 5.4

Radxa also offers configurations that include 1 GB RAM and 32 GB eMMC storage.

There is one reason to remain cautious.

A Pollen design document describes the Radxa Zero 3 target as provisional during development, noting that it was the closest available board rather than necessarily the final hardware choice.

The best-supported interpretation is therefore:

Microduck is built around an RK3566 compute architecture, and Radxa Zero 3W hardware has been used on real Microduck development units. Whether every final production robot uses an unchanged retail Zero 3W board remains unconfirmed.

That difference could matter significantly for cost.

Once production volume becomes large enough, moving from a commercial SBC toward a customized compute carrier or production-specific board can remove unused interfaces and reduce BOM cost.

Microduck’s “LiDAR” Is Actually an 8×8 ToF Sensor

Microduck is often described as having LiDAR.

Technically, developers should not imagine a rotating laser scanner such as those used on autonomous vehicles or mobile warehouse robots.

The robot uses a compact solid-state multi-zone Time-of-Flight depth sensor.

Pollen’s engineering documentation identifies an 8×8 depth matrix and explicitly displays the VL53L8CX sensor in its monitoring tools.

The VL53L8CX is manufactured by STMicroelectronics.

ST specifies the sensor as an 8×8 multi-zone ToF device capable of producing 64 depth zones and ranging up to approximately four meters. It also supports a wide field of view and frame rates of up to 60 Hz.

This is a useful example of how consumer BOM calculations can become distorted.

A developer might buy a complete VL53L8CX evaluation or breakout board.

A manufacturer does not necessarily do that.

At volume, the underlying sensor can be integrated directly into a custom PCB with shared power, communications, connectors, and mechanical mounting.

That eliminates:

• breakout-board PCB cost

• development connectors

• redundant regulators

• retail packaging

• distributor markup

The difference between “the development module costs $30” and “the production sensor costs a fraction of that” is one of the major reasons hobbyist replication costs cannot be directly compared with factory BOM costs.

Camera, IMUs, NFC and Audio: Where the BOM Becomes Less Certain

Pollen confirms that Microduck includes:

• a front camera

• two IMUs

• a microphone

• a speaker

• NFC hardware

The exact production part numbers are less clear.

Some third-party reverse engineering has proposed specific Bosch IMUs, camera sensors, NFC controllers and audio components.

Those should not yet be treated as established production facts.

For example, a technically plausible BOM could include a Bosch BMI-series IMU or an OmniVision/Sony camera sensor, but “plausible” is not equivalent to “confirmed.”

A credible hardware investigation should preserve uncertainty when evidence is incomplete.

For now, the following production details remain unresolved publicly:

• exact head IMU model

• exact body IMU model

• camera sensor manufacturer

• camera module supplier

• NFC controller

• audio codec

• microphone part number

• speaker part number

These components matter to the BOM, but they are unlikely to dominate it in the way the actuator system does.

Why the NP-F550 Battery Is a Smart Choice

Microduck uses a removable 2600 mAh NP-F550-format battery.

This is a mature battery format originating from photography and video equipment.

From a product-engineering perspective, this choice is highly rational.

A proprietary battery would force Pollen Robotics to manage:

• custom battery-pack development

• dedicated replacement inventory

• proprietary charging hardware

• long-term sourcing risk

• additional tooling

• aftermarket availability

The NP-F550 ecosystem already solves much of that problem.

Multiple manufacturers produce compatible batteries, chargers, holders, and accessories worldwide.

For users, replacement batteries are easy to obtain.

For Pollen, the mature supply chain likely lowers procurement and support complexity.

This type of decision is easy to overlook in a teardown, but standardized commodity components can have a significant impact on the economics of low-cost robotics.

So What Does Microduck Actually Cost to Build?

The exact answer is not publicly known.

A production BOM can only be established conclusively with access to Pollen Robotics’ supplier data or contract manufacturer records.

The most useful approach is therefore to build cost scenarios.

The largest unknown is the contracted price of the 15 XL330 actuators.

Consider three illustrative scenarios:

Scenario A: Aggressive Volume Procurement

Actuator cost: $10 each

Scenario B: Moderate Volume Procurement

Actuator cost: $14 each

Scenario C: Conservative Procurement

Actuator cost: $18 each

Using those assumptions:

Cost Category	Aggressive	Moderate	Conservative
15 XL330 actuators	$150	$210	$270
Compute, RAM and storage	$18–25	$22–30	$28–35
Camera, ToF and dual IMUs	$10–16	$15–22	$20–28
NFC, audio and related electronics	$5–8	$8–12	$10–15
Battery	$6–9	$8–12	$10–15
PCB, power system, cables and connectors	$10–15	$15–22	$20–30
Plastics, brackets, joints and fasteners	$20–30	$25–40	$35–50
Controller and USB-C accessories	$8–12	$10–15	$12–20
Estimated Direct Material Cost	$227–265	$313–363	$405–463

These figures are scenario estimates, not leaked or official Pollen Robotics BOM figures.

They deliberately exclude several costs that matter to a commercial product:

• assembly labor

• manufacturing testing

• yield loss

• tooling amortization

• packaging

• international freight

• warehousing

• warranty

• customer support

• payment processing

• software development

• engineering

• marketing

• taxes

• channel margin

The table nevertheless reveals something important.

The actuator contract is probably the single variable that determines whether the economics of the $399 Microduck look extraordinary or almost impossible.

If Pollen can obtain XL330-class actuators at a deeply discounted volume price while reducing the rest of the robot through custom integration, a direct BOM in the low-to-mid-$200 range becomes technically plausible.

If actuator costs remain close to ordinary distribution pricing, the economics become much more difficult.

There is currently no public evidence proving either case.

Why $399 May Be a Platform Price, Not Just a Hardware Price

The most interesting interpretation of Microduck emerges when its price is examined through software-platform economics.

Pollen Robotics is now part of Hugging Face.

Hugging Face’s core business has historically benefited from a familiar technology flywheel:

more developers

→ more models

→ more datasets

→ more tools

→ more community activity

→ greater platform value

Physical AI creates the possibility of extending that logic from software into machines.

Microduck can support a similar cycle:

More affordable robots

→ more developers own real hardware

→ more reinforcement-learning experiments

→ more trained policies and behaviors

→ more simulation and real-world data

→ better development tools

→ larger robotics community

→ more demand for compatible hardware

This changes the strategic meaning of hardware margin.

For a conventional consumer electronics company, maximizing gross margin per device is usually critical.

For a developer-platform company, the number of active machines in the hands of developers can have substantial strategic value of its own.

At $399, Microduck becomes accessible to a much wider population:

• robotics researchers

• university laboratories

• AI developers

• students

• reinforcement-learning researchers

• independent makers

• startups

• robotics enthusiasts

A comparable system costing $2,000 or $5,000 would produce a very different developer ecosystem.

Small Robots May Solve a Big Physical AI Problem

One of the hardest problems in embodied intelligence is collecting real-world experience.

Software models can be trained and tested millions of times inside digital environments.

Robots have to interact with physics.

That introduces friction:

• motors wear

• batteries discharge

• robots fall

• components break

• environments need resetting

• safety becomes relevant

• human supervision may be required

The larger the robot, the more expensive these failures become.

A full-size humanoid falling onto the floor may require:

• human intervention

• protective equipment

• repair

• replacement components

• safety procedures

Microduck weighs less than 800 grams.

That changes the experimentation equation.

A small robot can fall repeatedly during reinforcement-learning experiments at dramatically lower physical and financial risk.

This is one of Microduck’s strongest strategic advantages.

Its contribution to Physical AI may therefore come less from achieving human-scale capability and more from making real-world learning iterations cheap enough to run frequently.

The China Manufacturing Connection

Microduck also illustrates how modern robotics increasingly combines software developed across the United States and Europe with a highly mature Asian electronics supply chain.

Its main processor comes from Rockchip, a Chinese semiconductor company.

Radxa, whose Zero 3W platform appears throughout Pollen’s Microduck development documentation, operates from Shenzhen.

A product in Microduck’s price range is also particularly well suited to the manufacturing ecosystem around Shenzhen and the Greater Bay Area, where suppliers can provide:

• PCBs

• SMT assembly

• camera modules

• battery packs

• injection-molded plastics

• metal brackets

• wiring harnesses

• connectors

• rapid tooling

• final assembly

The economics matter.

A robotics startup trying to recreate this entire supply base independently in Europe or North America would face significantly different tooling, sourcing, lead-time, and unit-cost constraints.

The broader lesson extends beyond Microduck:

Modern robotics is increasingly becoming a combination of global AI software, specialized actuator technology, commodity semiconductor platforms, and highly optimized Asian manufacturing.

Companies that can integrate all four layers will have a structural cost advantage.

Is Microduck Open Source?

This requires a precise answer.

Microduck’s software stack is open source. Its complete production hardware is not.

Pollen Robotics explicitly asks publications not to describe Microduck as open-source hardware.

Its official legal and press materials state that the mechanical and electronic design files are not released under an open-source license unless specifically stated otherwise.

The software side is much more open.

Developers can work with code covering areas such as:

• robot control

• reinforcement learning

• policy deployment

• simulation

• runtime services

• developer tooling

Simulation and 3D assets are also available in the broader Microduck development ecosystem, although individual assets may carry different license conditions.

This creates an important distinction for commercial developers.

An Apache-licensed software repository does not automatically grant rights to reproduce:

• Microduck’s industrial design

• Pollen Robotics trademarks

• Hugging Face trademarks

• unreleased electronics

• proprietary mechanical files

• assets distributed under noncommercial licenses

Developers can build software around Microduck and study the architecture.

A commercial hardware clone requires a separate intellectual-property analysis.

What Is Confirmed and What Is Still Unknown?

Publicly Confirmed

Microduck has:

• 15 degrees of freedom

• Rockchip RK3566 compute

• 1 GB RAM

• 32 GB storage

• a front-facing camera

• 8×8 depth sensing

• two IMUs

• microphone and speaker

• NFC

• Wi-Fi and Bluetooth

• NP-F550-format battery

• approximately 25 cm height

• weight below 800 g

• onboard 50 Hz policy execution

• $399 introductory pricing

Strongly Supported by Engineering Documentation

Available Pollen documentation strongly supports:

• DYNAMIXEL XL330-series actuators

• 15 servo devices

• Radxa Zero 3W use on real Microduck development hardware

• STMicroelectronics VL53L8CX depth sensing

Highly Likely but Not Fully Confirmed

• XL330-M288-T as the exact actuator variant

Still Unknown Publicly

• Pollen’s actuator contract price

• exact production BOM

• final production motherboard design

• exact camera sensor

• exact IMU models

• exact NFC controller

• exact audio codec

• complete supplier list

• manufacturing labor cost

• unit gross margin

Why Microduck Matters

The easy story about Microduck is:

“Fifteen servos cost more than the robot.”

The more important story is what that pricing could do to the robotics market.

For years, sophisticated robotics experimentation has been constrained by hardware cost.

AI software became widely accessible much faster than physical machines did.

Open-source models can be downloaded.

Cloud GPUs can be rented.

Simulation environments are available to almost anyone.

Physical robots remain expensive.

Microduck attacks that mismatch.

For $399, developers gain access to a physical system that combines:

• sensing

• actuation

• onboard compute

• reinforcement learning

• simulation

• sim-to-real deployment

• programmable behavior

That brings a meaningful Physical AI experimentation loop into a price category closer to consumer electronics than laboratory robotics.

The long-term significance may have little to do with robotic ducks specifically.

The more consequential possibility is that low-cost standardized robots could become the physical equivalent of development boards.

Raspberry Pi and Arduino helped reduce the cost of experimenting with computing and electronics.

Affordable AI robots could perform a similar role for embodied intelligence.

If that happens, competition will shift.

The winning platforms may not simply be those with the strongest motors or most sophisticated mechanical designs.

They may be the platforms with the largest combination of:

• installed hardware

• trained policies

• simulation environments

• datasets

• developer tools

• reusable behaviors

• community contributions

That is why the $399 price matters.

Microduck may ultimately be remembered less for being an inexpensive biped robot and more for testing whether Physical AI can develop a mass developer ecosystem similar to the ecosystems that transformed software, mobile computing, and open-source AI.

Frequently Asked Questions

How much does Microduck cost?

Microduck launched at an introductory price of $399 before taxes and shipping. Pollen Robotics opened pre-orders on August 27, 2026.

How many motors does Microduck have?

Microduck has 15 degrees of freedom. Pollen’s control architecture also describes 15 servos operating within the robot.

What servos does Microduck use?

Public engineering evidence strongly points to the ROBOTIS DYNAMIXEL XL330 family. The XL330-M288-T appears to be the most likely exact variant, although the complete production part number has not been formally listed in Pollen’s main consumer specification.

How much would 15 XL330-M288-T servos cost?

At ROBOTIS U.S. retail pricing of $27.49 each, 15 units cost approximately $412.35, exceeding Microduck’s $399 introductory price.

What processor does Microduck use?

Microduck officially uses a Rockchip RK3566, paired with 1 GB RAM and 32 GB storage.

Does Microduck use a Radxa Zero 3W?

Pollen Robotics has demonstrated Microduck software running on real Radxa Zero 3W hardware, and its engineering documentation repeatedly references the board. The RK3566 architecture is confirmed, although Pollen has not established that every production robot necessarily contains an unchanged retail Zero 3W.

What depth sensor does Microduck use?

Pollen engineering documentation identifies the STMicroelectronics VL53L8CX, an 8×8 multi-zone Time-of-Flight sensor capable of generating 64 depth zones.

What is Microduck’s estimated BOM cost?

Pollen Robotics has not published an official BOM.

Depending on actuator pricing and hardware integration, a reasonable scenario model can produce direct-material estimates ranging from roughly the mid-$200s to above $400.

The biggest unknown is Pollen’s real procurement price for its 15 actuators.

Is Microduck open-source hardware?

No.

Pollen Robotics states that Microduck’s open-source designation applies to its software stack. Complete mechanical and electronic design files have not been released as open-source hardware.

Why is Microduck important for Physical AI?

Microduck combines real-world sensing, bipedal actuation, reinforcement learning, onboard policy execution, simulation, and sim-to-real workflows at a $399 introductory price.

That significantly lowers the cost of conducting Physical AI experiments on real hardware.

Research Methodology and Sources

This analysis uses publicly available information published or maintained by:

• Pollen Robotics — Microduck official specifications, press materials, product information and engineering documentation

• Pollen Robotics GitHub repositories — runtime architecture, hardware bring-up documentation, control-loop documentation and ToF implementation

• ROBOTIS — DYNAMIXEL XL330 product specifications and U.S. retail pricing

• STMicroelectronics — VL53L8CX technical specifications

• Radxa — Zero 3W hardware specifications and RK3566 platform information

Cost estimates in this article are analytical scenarios based on publicly observable component economics. They are not Pollen Robotics supplier quotations or disclosed manufacturing costs.

Information was reviewed and updated on September 4, 2026.