MicroDuck, Taken Apart: The Hardware Behind a Tiny Walking Robot
An illustrated tour of the shells, sensors, actuators and linkages inside Pollen Robotics’ small biped.
MicroDuck has a surprisingly expressive silhouette: a large camera eye, a broad beak, a slender neck and two substantial feet. Even standing still, it looks ready to investigate something on your desk.
That character comes from a physical arrangement of parts. The head needs support and movement. The legs need to carry it. The feet need somewhere to meet the ground. Computing, sensing and power must fit around all those moving connections.
Our 2-minute, 56-second video explores that arrangement by separating MicroDuck into its published components, then bringing everything back together.
One distinction matters from the start: this is an exploded visualization of the public simulation assembly, not a filmed disassembly of a physical robot. The part shapes and scale come from Pollen Robotics’ released geometry. The model does not provide a complete manufacturing bill of materials or every detail of the current electronics.
That still makes it a rich engineering reference. Here is what becomes visible when we give its components room to breathe.
First, separate product dimensions from model measurements
Pollen’s official press kit describes MicroDuck as approximately 25 centimetres tall and 14 centimetres wide.
The published model, measured in the standing pose used for this visualization, occupies a different bounding box: 271.62 mm high, 142.06 mm wide and 122.69 mm deep.
These numbers answer different questions. The product figures give nominal dimensions. The model figures describe the outermost geometry in one particular posture. Joint angles can change that envelope, but the available sources do not establish the full reason for the difference. The model measurement should not replace the product specification.
Individual part dimensions need similar care. A mesh measurement records the size of a published shape; it does not supply manufacturing tolerances, a material specification or a guaranteed fit on a production unit. Throughout the video, those measurements are identified as source-model dimensions.
The shells establish the silhouette
The first pieces to move away are the head dome, head cover, torso panels and thigh covers. Their separation reveals how much of MicroDuck’s appearance comes from a few relatively large surfaces surrounding a compact mechanism.
The curved dome gives the head its recognizable outline. The torso panels gather the central structure into a rounded body. The thigh covers soften the visual transition from body to exposed leg hardware.
Removing those surfaces also changes how you read the robot. A friendly outline becomes a set of mounting points, brackets and mechanical connections.
Color helps preserve the relationship between those two views. The visualization uses the official Cream shell reference, orange beak and foot structures, and yellow eye and sole accents. Internal pieces retain their separate source colors, making it easier to follow a component as it moves away from the assembly.
Inside the head: optics and a supporting face
The head view separates the front face, eye insert, lens and M12 lens holder. The eye’s appearance depends on several pieces sharing a position and orientation, rather than one decorative disk.
The lens needs a holder, and that holder needs a mounting relationship with the surrounding structure. Exploding these pieces makes that relationship much easier to see than looking through the assembled face.
The current product specification includes a front camera and an 8×8 time-of-flight depth sensor. A time-of-flight sensor estimates distance using the travel time of light. Here, the useful distinction is between recognizing what is in front of the robot and measuring how far away it is.
The public meshes are not a complete, component-level model of both sensor systems. The film labels their product functions without inventing missing circuit boards or chips.
The beak gives the face a practical job
The beak is also a gripper. Its jaw and contact inserts provide a small mechanism for picking up lightweight objects.
Separating the rigid jaw from the soft pads reveals the different jobs of shape and contact. The rigid component establishes the form; the softer surfaces provide the interface with an object.
The published model contains distinct jaw and upper-mouth inserts, which we retain as separate pieces. Their names and appearance support that explanation, but do not establish a precise polymer grade or production hardness.
There is also a motion limitation to respect: the jaw is fixed in the simulation assembly used here. The exploded movement illustrates the component arrangement. It is not a simulation of the complete grasping mechanism operating under load.
The neck stacks several directions of movement
Below the head, a small collection of carriers, bearings and links forms an articulated neck.
The names describe the motion: pitch tilts, yaw turns from side to side, and roll changes the head’s sideways lean. Separating the corresponding carriers reveals how those directions are arranged through the structure.
This is a useful way to understand an expressive robot. A head pose that looks like one gesture can depend on several joints working together.
Get Gnana Prasath’s stories in your inbox
Join Medium for free to get updates from this writer.
It also explains why the neck deserves a close-up. These compact components connect a relatively large head to the body while providing several orientations. Their arrangement is central to both MicroDuck’s mechanics and its character.
Electronics: distinguish the visible model from today’s hardware
The electronics chapter brings out the modeled compute-board envelope, Robot HAT, speaker, battery envelope and supports.
This is where a simulation model’s history becomes particularly important. One mesh is named pcb__raspberry_pi_zero_2_w; the battery mesh is named np_f970. Those filenames should not be treated as current purchasing instructions.
The official product specification identifies RK3566-based computing and a removable NP-F550 battery. The released runtime also provides evidence of the Radxa Zero 3W platform. The video therefore identifies the visible board and battery shapes as older model representations.
There is a useful additional source for electronics enthusiasts: Pollen publishes an RPI Robot HAT hardware project, including KiCad files and production outputs. Its functions include motor connections, an IMU and audio support.
That board is a more detailed engineering reference than the simplified simulation PCB. It still does not make the simulation assembly a complete, verified model of every production electronic component. The product specification also lists head and body IMUs, microphones, and NFC antennas in the head and beak. An IMU measures motion and orientation-related signals; these additional systems are not all represented by complete, separate meshes here.
Fifteen actuators, with different responsibilities
The public assembly contains 15 instances of the Dynamixel XL330 servo mesh. Seeing them apart makes their distribution through the body, neck and legs much clearer.
A servo combines a motor with mechanisms for controlling its output position. In this teardown, each servo remains a complete unit. We do not add imaginary internal gears, motors or electronics to a mesh that represents the housing.
The count also needs context. The model exposes 14 actuated joints in its walking control arrangement. The real robot’s mouth is controlled separately, as described in the official runtime’s joint and mouth-control tables.
That distinction explains why “15 motors” and “14 policy-controlled joints” can both be correct.
Hips, legs and feet turn articulation into support
At the hips, the exploded view exposes couplers and bearings before following the structure through the thighs, rigidity plates and lower-leg links.
These pieces are easier to understand as a chain. Motion introduced near the body must pass through connected structures toward the ankle. Bearings support rotation; links and brackets establish the relationships between those rotating parts.
The model contains separate left and right versions of covers, ankles and feet. Keeping those identities visible helps explain the mirrored assembly; the mesh names alone do not establish whether any production parts are interchangeable.
At ground level, the foot separates into an ankle carrier, foot body and sole. The broad sole provides the visible contact surface, while the components above it connect that surface to the leg.
The foot-body mesh measures approximately 40.09 × 54.00 × 16.91 mm along its source axes. That describes the modeled foot body, not the entire assembled foot or its contact area.
An exploded assembly is not automatically a full parts list
The complete view contains 38 distinct mesh designs and 70 placed visual instances. Repeated parts, such as servos and bearings, account for the difference.
Those counts deliberately exclude duplicate collision geometry. Simulation software can use an additional shape to calculate contact without representing an additional physical object.
The counts also do not include a complete inventory of fasteners, cable lengths, battery contacts or motor internals. Those omissions set boundaries on what the visualization can explain.
How the visualization preserves the engineering
The geometry comes from a pinned official MicroDuck model. Its assembly description and 38 STL files were checked against that source version.
The parts keep their original vertex positions and assembly relationships. Exploded views separate them by translation; camera movement makes small details larger on screen. Individual components are not resized to make a shot more dramatic.
Surface normals — the directions used to calculate lighting — were reconstructed to display the geometry cleanly. This improves how a surface catches light without changing its shape.
What stays with you after reassembly
MicroDuck becomes easier to understand when its personality can be traced back to physical decisions: a supported eye, a useful beak, stacked neck joints and broad feet.
The broader lesson is to read public robot models with both curiosity and precision. They can reveal a great deal about arrangement and movement, while leaving important manufacturing questions unanswered.
Credit belongs to Pollen Robotics and Hugging Face for the robot and published assets. The repository identifies the public geometry used here under Creative Commons BY-NC-SA and software under Apache-2.0. These released meshes do not constitute a complete open-source hardware design; Pollen’s product-level open-source statement covers software. Consult the linked sources for their terms. The Awesome MicroDuck community index is a useful starting point for further exploration.
Watch the video at the top to follow the complete separation and reassembly. Which part would you like to explore more closely: the grasping beak, the articulated neck or the leg mechanism? Leave a comment with the next detail you want unpacked.
