Microduck Hardware Explained: How to Build an Original DIY RL Biped Robot

Microduck has quickly become one of the most talked-about small robots of 2026. Pollen Robotics presents it as a 25 cm, roughly 800 g biped with 15 actuators, a camera, depth sensing, two IMUs, and a 50 Hz onboard reinforcement-learning control loop. Its introductory preorder price is $399, while its software, simulator, and reinforcement-learning workflow are published for developers to inspect and extend.

That combination naturally raises a question:

> Can makers build a Microduck-style reinforcement-learning robot from commercially available parts?

The short answer is yes, if the goal is an original DIY biped that explores the same engineering ideas. It is not responsible to promise a cheap, one-to-one Microduck clone. Pollen Robotics has released the software stack, but it has not published a complete official hardware bill of materials, production-ready CAD, PCB schematics, cable drawings, or assembly instructions.

This guide therefore does three things carefully:

1. separates official specifications from facts found in the official source code;
2. identifies practical development components without presenting them as an official BOM;
3. proposes an original DIY robot architecture that does not require copying Microduck's shell, branding, production files, or unpublished electronics.

Important independence notice

This is an independent technical commentary and parts-selection guide by rcdrone.top. rcdrone.top is not affiliated with, authorized by, sponsored by, or endorsed by Pollen Robotics or Hugging Face. “Microduck,” “Pollen Robotics,” “Hugging Face,” “DYNAMIXEL,” “ROBOTIS,” and all other product names and marks belong to their respective owners and are used only to identify and discuss the relevant products.

No rcdrone.top product should be marketed as an official Microduck part unless it is an authentic part sold under the manufacturer's own brand and accurately identified as such. Third-party boards and kits should use names such as RL Biped Development Board, 15-Servo Robot Wiring Kit, or Microduck Software Development Compatible, followed by a clear compatibility qualification.

What Pollen Robotics officially confirms

The official Microduck product page confirms the following high-level specifications:

These facts are useful, but they are not a component-level BOM. Exact part numbers must be treated separately.

What can be established from the official software repository

A robot runtime has to know which serial port to open, which devices to address, which camera overlay to load, and which registers to read. For that reason, the official software repository reveals more about the electrical architecture than the marketing page.

The current public code and design documentation point to this architecture:

These are source-code-supported findings, not a Pollen Robotics-issued production BOM. Hardware may change before or during production.

The most important hardware choice: 15 smart servos

Microduck is not built around ordinary hobby PWM servos. Its control system depends on a multi-drop digital servo bus and joint-state feedback.

The evidence points to 15 ROBOTIS DYNAMIXEL XL330-M288-T-class actuators. Fourteen joints participate in the learned policy; the fifteenth actuator operates the mouth or beak under higher-level control.

XL330-M288-T reference specifications

For a learned biped, the 18 g mass is just as important as the torque figure. Replacing each actuator with a 55 g serial servo adds approximately 555 g across 15 joints before brackets become larger. That changes limb inertia, center of mass, impact behavior, and the simulation-to-real transfer problem.

Why a high-torque hobby servo is not automatically a substitute

A candidate actuator should be evaluated against all of these requirements:

• approximately 20 × 34 × 26 mm packaging;
• a target mass near or below 20 g;
• low backlash and repeatable zero position;
• absolute joint position feedback;
• velocity and preferably current or torque-related feedback;
• temperature and supply-voltage telemetry;
• deterministic multi-servo bus access;
• enough bandwidth for a 50 Hz outer control loop;
• a documented host protocol;
• a supply system that can survive synchronized acceleration and falls.

Comparing stall torque alone is misleading. Stall torque is a short-duration limit, not a safe continuous operating point.

Servo alternatives: what is realistic?

Option 1: ROBOTIS XL330-M288-T — the lowest software risk

Using the same actuator family gives a DIY developer the best chance of matching the published bus behavior, register map, feedback signals, dimensions, and learned dynamics.

The disadvantage is cost. At normal retail pricing, 15 genuine XL330 servos can cost as much as or more than the introductory price of the complete robot. The value of this route is compatibility and development access, not necessarily a lower total price.

Best for: developers who want to study the official runtime, joint control, and sim-to-real training with the fewest actuator changes.

Option 2: Unitree S288 — a promising mechanical-size alternative

The Unitree S288 is unusually interesting because its external dimensions are close to the XL330 class and its listed weight is approximately 19.5 g. It uses a brushless motor, dual absolute encoders, and a digital half-duplex bus.

It is not electrically or software compatible with the XL330:

• recommended supply is in the 12.6 V class rather than 5 V;
• communication is Unitree's fixed 6 Mbps protocol rather than DYNAMIXEL Protocol 2.0;
• register layout and control semantics are different;
• joint dynamics and controller gains differ;
• the motor driver and robot policy must be adapted and validated.

The S288 is best described as a candidate for a new RL biped design, not a drop-in Microduck replacement.

Best for: an original high-performance robot with a rewritten driver, separate power architecture, and retrained policies.

Option 3: Feetech STS3215 — affordable torque, different robot

The STS3215 offers a digital serial bus, position feedback, and substantially more torque. Typical versions are approximately 45.2 × 24.7 × 35 mm and 55 g.

Those numbers make it unsuitable as a one-to-one replacement in a 700–800 g robot. A robot designed around STS3215 servos will be larger and heavier and will require original brackets, a new mass model, new power distribution, new control software, and new RL training.

It can still be a good teaching actuator for a larger tabletop biped.

Best for: a lower-cost, larger educational robot where exact Microduck geometry and policies are not required.

Option 4: ordinary PWM servos — suitable only for a simplified project

PWM servos can move a robot through scripted poses, but most models do not provide the synchronized position, velocity, current, voltage, and temperature telemetry expected by this architecture. They are not recommended for reproducing the published learned-control loop.

Best for: slow pose demonstrations, animatronics, or an introductory duck-inspired robot with no claim of policy compatibility.

New compact alternative: Feetech HD-1910-C001 for a Microduck-style DIY build

For builders developing their own small Microduck-style biped, the Feetech HD-1910-C001 coreless metal gear servo is a more suitable candidate to evaluate than the larger STS3215 when compact joint packaging is the priority. Its listed 34 × 20 mm body footprint, dual-shaft structure, and multiple mounting points make it a more targeted starting point for small articulated mechanisms. The STS3215 remains a useful option for the larger educational robots described above.

The HD-1910-C001 combines a coreless motor and metal gears with a 3-pin TTL serial-bus interface. Its listed operating range is 5–8.4 V, with 10 kg.cm stall torque. That torque figure is for stall conditions, not sustained joint loading; actual power supply sizing and usable joint performance should be established during bench testing.

Choose it as an actuator for an original DIY design, rather than assuming it is a plug-in replacement for the official XL330 setup. Check the mounting drawing, actual servo mass, travel range, bus protocol, feedback support, and controller settings before committing to a full robot. A TTL connector alone does not establish DYNAMIXEL Protocol 2.0 compatibility, and a different actuator can require controller, calibration, and robot-model changes.

Best for: compact Microduck-inspired joints and small biped prototypes where the builder can adapt the mechanics and control software. Select the HD-1910-C001 Servo option on the product page; the URT2 Test Board is a separate purchase and is not included with the servo.

Main computer and onboard AI

The official software points to the Radxa Zero 3W family. It uses the Rockchip RK3566, a quad-core Cortex-A55 SoC with Mali-G52 graphics, hardware video support, and a small NPU. The board also provides Wi-Fi and Bluetooth in a Raspberry Pi Zero-like 65 × 30 mm form factor.

That choice makes sense because the onboard computer may need to handle several tasks simultaneously:

• ONNX policy inference at 50 Hz;
• joint feedback and safety logic;
• 720p video streaming through WebRTC;
• Rockchip hardware H.264 encoding;
• low-rate RKNN object detection;
• Bluetooth gamepad and phone communication;
• Wi-Fi provisioning;
• signed software updates and rollback.

Can a Raspberry Pi Zero 2 W be used instead?

It can be useful for a motion-only or simplified development robot, and official notes indicate that an earlier prototype used a Pi Zero 2 W. It is not a transparent replacement for the full current stack. Rockchip MPP video encoding, RKNN inference, device-tree overlays, serial mappings, and deployment scripts would need to be changed.

For a first product offering, the safer parts strategy is:

• Radxa Zero 3W with a clearly stated RAM/eMMC configuration;
• compatible heat spreader;
• 40-pin header;
• verified IMX219 ribbon cable;
• a tested operating-system image and version-specific setup guide.

Do not market another SBC as “fully Microduck compatible” without testing the complete software and peripheral stack.

Orientation sensing: an IMU on the servo bus

One of the most elegant parts of the architecture is the imu_to_dxl board. Instead of making the Linux computer read the main IMU over a separate I²C transaction, the board behaves like another device on the DYNAMIXEL bus.

The official control design uses bus ID 200 and reads a compact sensor block alongside the 15 actuators. The evidence points to an ST LSM6DSV16X, a six-axis IMU with an onboard SFLP sensor-fusion engine.

The control loop consumes:

• three gyroscope axes, configured around ±500 degrees per second;
• three quaternion components in compact half-precision form;
• a reconstructed fourth quaternion component on the host.

A legally cleaner DIY alternative

Rather than copying an unpublished Pollen PCB layout, design an original interface board from public datasheets and protocol documentation:

• LSM6DSV16X;
• a small MCU such as an STM32G0 or CH32V203;
• an original 3.3 V TTL half-duplex circuit;
• an original PCB layout and connector placement;
• independently written firmware implementing the documented host-facing behavior.

If the firmware reuses Apache-2.0 source code, preserve the copyright, license, and NOTICE obligations. If it is independently implemented from public protocol specifications, document that process and keep design records.

This type of board is a strong future product opportunity for rcdrone.top, but it should be named generically—for example, IMU-to-Digital-Servo Bus Adapter—and its compatibility statement should identify tested software versions.

Depth sensing: VL53L5CX and VL53L8CX

The robot uses an ST multi-zone time-of-flight sensor rather than a scanning mechanical LiDAR. The VL53L5CX and newer VL53L8CX families can produce an 8 × 8 distance grid over I²C. Public software evidence points to address 0x29 and an operating rate around 15 Hz in this application.

For a DIY project, a stock VL53L8CX breakout with a Qwiic or STEMMA QT connector is the simplest choice. It avoids copying the original sensor PCB and is easy to replace.

Potential product variants include:

• bare VL53L8CX breakout;
• breakout plus short Qwiic cable;
• breakout plus mounting hardware;
• tested sensor-and-SBC bundle.

Sensor field of view, window material, mounting angle, crosstalk calibration, and I²C pull-ups should be documented. A module using the same chip is not automatically mechanically calibrated for the same robot.

Camera: IMX219 remains the practical choice

The software configuration points to an IMX219 module, commonly sold as the Raspberry Pi Camera v2. The published media path uses a 1080p30 sensor mode and produces a 720p30 WebRTC stream through the Rockchip hardware encoder.

An IMX219 module is an attractive product because it is widely available and supported by many SBCs. However, buyers need to know:

• ribbon-cable pitch and contact orientation;
• standard, NoIR, or wide-angle lens type;
• horizontal field of view;
• whether the image must be rotated in hardware;
• which device-tree overlay was tested;
• whether the module works with hardware H.264 encoding.

Wide-angle and NoIR versions should be sold as development alternatives, not as exact optical replacements for an unverified production camera.

Audio options

The official device-tree configuration identifies a TLV320AIC3104 audio codec. The production HAT itself is not open hardware, and its complete schematic is not available.

An original DIY robot can use one of three approaches:

1. USB audio adapter: easiest to integrate, but consumes space and a USB connection.
2. Original I²S codec board: supports microphone input and speaker output, but needs more Linux and PCB work.
3. MAX98357A I²S amplifier: compact and easy for sound output, but does not provide microphone recording.

For an early rcdrone.top development kit, audio should be optional. A motion-and-sensing kit is easier to support than a board that combines high-current power, half-duplex servo communications, camera, microphone, speaker, and codec clocks in its first revision.

Power design: do not copy an unresolved assumption

This is the area where a responsible DIY guide must be conservative.

Public reverse-engineering work associates the robot with an NP-F-series 2S lithium-ion battery and reports a 6.6–8.2 V working range. A retail XL330-M288-T, however, is officially rated for only 3.7–6.0 V.

Until a shipping production robot is measured, do not assume that an NP-F battery is connected directly to retail XL330 servos. The final product may contain an unpublished regulator, separate power rails, a customized actuator variant, or another detail not represented in public material.

Safer DIY power architecture

For a robot using retail XL330-M288-T servos:

• regulate the servo rail within the manufacturer's published voltage range;
• provide a separate, stable 5 V rail for the SBC when practical;
• use a fuse or resettable protection device;
• include a physical power switch and accessible emergency stop;
• size wiring for transient current and voltage drop;
• do not design from “15 × stall current” as a continuous operating condition, but do test worst-case synchronized transients;
• log rail voltage during stand-up, impacts, and blocked-joint events;
• use a protected battery pack and a charger intended for that chemistry and series count.

Any future rcdrone.top power kit should state input range, regulated outputs, continuous and peak ratings, connector polarity, wire gauge, protection behavior, and the exact servo/SBC combination tested.

The software stack that makes the robot interesting

The public robot runtime is primarily written in Rust and deliberately avoids a large robotics framework. Its services are separated by responsibility:

The processes communicate through a JSON-RPC contract over Unix sockets. This separation is valuable for a DIY robot: a failed behavior process does not have to destroy networking or the update path.

Reinforcement learning and sim-to-real

The public workflow uses MuJoCo simulation, PPO-style reinforcement learning, domain randomization, and ONNX export. Current documentation describes a 50 Hz policy interface with 61 observation values and 14 joint actions.

The crucial lesson is that a trained policy contains assumptions about the physical robot:

• link mass and inertia;
• center of mass;
• servo speed and torque response;
• backlash and friction;
• command latency;
• controller gains;
• supply-voltage behavior;
• joint limits;
• ground contact and foot geometry.

Changing to another servo or changing the frame is not merely a mechanical substitution. The simulator must be updated, and the policy will usually require retraining and real-world validation.

For safety, first tests should use a support rig, reduced gains, conservative joint limits, a clear floor area, eye protection where appropriate, and an accessible power disconnect. Small robots can still pinch fingers, damage gears, or eject broken printed parts.

A practical three-stage DIY roadmap

Stage 1: Sensor and software bench kit

Start without legs.

Suggested parts:

• Radxa Zero 3W;
• IMX219 camera;
• VL53L8CX breakout;
• LSM6DSV16X breakout;
• Bluetooth gamepad;
• ribbon cables, Qwiic cables, and a stable bench supply.

This stage validates Linux, camera capture, depth frames, orientation data, Bluetooth, and software deployment without the cost and risk of 15 actuators.

Stage 2: Five-servo single-leg development rig

Suggested parts:

• five XL330-M288-T servos;
• genuine or verified TTL servo interface;
• 5 V high-current regulated supply;
• fuse, switch, and safe wiring;
• an original test frame;
• printed guards and a support fixture.

This stage verifies bus timing, joint IDs, calibration, current limits, position feedback, gain settings, and simulation parameters. A single-leg kit is also a more accessible product than forcing every reader to buy 15 servos at once.

Stage 3: Original 14+1 actuator biped

Only after the first two stages should a builder integrate:

• 10 leg actuators;
• four head and neck actuators;
• one optional beak or gripper actuator;
• original links and enclosure geometry;
• original sensor mounts;
• original wiring and power-distribution board;
• an original appearance that does not imitate Microduck's distinctive shell or colorways.

The simulator should use measured masses, joint axes, actuator limits, and inertias from the builder's own hardware—not copied production geometry.

Recommended parts categories for rcdrone.top

The strongest commercial opportunity is not an unverified “clone kit.” It is a set of well-documented development components that solve specific integration problems.

Products that can be introduced first

• Radxa Zero 3W boards with explicit RAM and storage variants;
• genuine ROBOTIS XL330-M288-T servos individually and in 3-, 5-, and 15-unit bundles;
• XL330 3-pin JST servo cables in multiple verified lengths;
• DYNAMIXEL-compatible TTL development interfaces;
• IMX219 camera modules and matching ribbon cables;
• VL53L8CX 8 × 8 ToF breakouts;
• LSM6DSV16X development boards;
• Bluetooth gamepads;
• protected batteries, battery holders, and compliant chargers;
• regulated 5 V power modules with honest continuous and peak ratings;
• fuses, switches, XT30/JST power harnesses, and M2 hardware;
• bearing and fastener development packs for original small bipeds.

Higher-value original products to develop

• IMU-to-digital-servo-bus adapter;
• high-current robot interface HAT with a documented TTL bus;
• 15-servo color-coded wiring harness;
• single-leg RL development frame;
• tested sensor-and-compute bundle;
• original biped frame designed from clean-sheet CAD;
• versioned software image and bring-up documentation.

Every compatibility-oriented listing should include a matrix like this:

Five mistakes to avoid

1. Calling a reconstructed list an “official Microduck BOM”

It is not official unless Pollen Robotics publishes it as such. Use evidence-based hardware analysis, source-code-supported component list, or development parts guide.

2. Selling simulation meshes as commercial production parts

Simulation geometry may be incomplete for manufacturing, and third-party reconstructions report non-commercial licensing on relevant mesh assets. Do not sell printed derivatives without confirming the license and obtaining any necessary permission. Create original CAD instead.

3. Using Microduck logos, product photography, or packaging artwork without permission

Use original diagrams and photographs of the parts you actually sell. Link to the official product page for reference instead of copying its visual assets.

4. Describing a third-party kit as official or endorsed

Use a prominent independence statement. Avoid confusing names such as “Official Microduck HAT,” “Microduck Replacement Board,” or a product logo that resembles the original branding.

5. Assuming matching dimensions mean matching behavior

An actuator with the same dimensions may use another voltage, protocol, feedback model, torque curve, and thermal limit. A new driver and policy validation may still be required.

Frequently asked questions

Is Microduck open source?

Its public robot software is licensed under Apache License 2.0, and Pollen Robotics also publishes the simulation and reinforcement-learning workflow. That does not automatically make every mechanical design, PCB, logo, photograph, or product name open for commercial reproduction.

Is this an official Microduck BOM?

No. It combines official published specifications with component identities supported by the official software configuration. It should be treated as a development guide, not a production BOM.

Can I build a cheaper Microduck with standard servos?

You can build a duck-inspired biped, but standard PWM servos do not reproduce the smart-servo feedback and deterministic digital bus used by the published control architecture. The result will require different electronics, software, mechanics, and training.

Can Unitree S288 servos replace XL330 servos?

They are promising for an original compact robot because the size and weight are close. They are not drop-in replacements: voltage, protocol, commands, feedback, and dynamics differ.

Can I use the official learned policies after changing the hardware?

Do not assume so. Even small changes in mass, inertia, friction, backlash, motor response, controller gains, or timing can break sim-to-real behavior. Update the simulation and retrain or retune the policy.

What is the best first purchase?

For software and perception work, start with the compute-and-sensor bench kit. For locomotion work, start with a five-servo single-leg rig. Both paths reduce risk and create reusable parts for a later full robot.

Final verdict

Microduck is valuable because it shows that a small reinforcement-learning robot can combine lightweight smart actuators, synchronized state feedback, onboard inference, vision, depth sensing, Bluetooth control, and robust software updates in one approachable platform.

The responsible DIY opportunity is not to copy its distinctive product design. It is to learn from the published architecture and build a new robot around:

• lightweight feedback servos;
• a deterministic half-duplex bus;
• a Linux SBC with hardware video support;
• a fused IMU, multi-zone ToF sensor, and compact camera;
• original mechanics and electronics;
• a measured MuJoCo model;
• retrained policies matched to the actual hardware.

That path gives makers a useful project, gives developers room to improve the design, and gives rcdrone.top a sustainable catalog of smart servos, sensors, compute boards, cables, power modules, and original development kits—without presenting a third-party product as the official robot.

Sources and further reading

• Microduck official product page
• Pollen Robotics Microduck runtime repository
• Microduck software architecture
• Microduck robot control design
• Microduck reinforcement-learning repository
• ROBOTIS XL330-M288-T specifications
• Unitree S288/J288 specifications
• Feetech STS3215 product information

---

Update log

September 11, 2026: Added the Feetech HD-1910-C001 as a compact servo candidate for Microduck-style DIY builds, including its key specifications, mechanical advantages over the larger STS3215 for compact joints, and integration considerations. Retained all previous servo options and recommendations for their respective use cases.