Last night, my WeChat Moments feed was flooded with a duck.

Not a plush toy, but a robot duck—25 cm tall, with 15 motors, able to walk, fall, and get back up, and packed inside with a Rockchip RK3566.

It's called Microduck, priced at $399, and its pre-orders surpassed $2.6 million within 24 hours.

The next day, Rockchip's share price hit a limit-up straight away.

My feed basically split into two camps: one group reposted saying "the iPhone moment for embodied intelligence has arrived," while the other said "consumer-electronics SoCs have finally found a new story."

My first reaction was—how on earth does this RK3566 actually run a duck?

Setting my cup of tea aside for a moment, today, following the old habit of an embedded engineer, let's take this robot duck apart and check each of its hardware requirements against the chip, to see how the RK3566 meets them one by one.

Microduck's Hardware Spec Sheet

Microduck is a creation of Pollen Robotics (the folks behind Hugging Face) and is highly open-source, with its press kit made publicly available. I've picked out a few key specs from its officially released data, because whether the RK3566 can handle it all comes down to these numbers:

l Size / Weight: 25 cm tall, under 800 g

l Actuators: 15 motors (bipedal legs + beak + tail + head)

l Sensing: 1 camera + one 8×8 ToF LiDAR + dual IMU

l Main controller: Rockchip RK3566

l Memory / Storage: 1 GB LPDDR4 + 32 GB eMMC

l Software: Apache 2.0 licensed open-source SDK; robot-side software written in Rust

l OTA: supports over-the-air policy updates (three independent services: robotd / updaterd / configd)

Just these specs alone make any embedded engineer sit up—this is no longer a "little toy." It's a small robot platform that runs Linux and can move and perceive.

With the price held at $399, the robot's SoC cost is squeezed to the point where basically only a mature AIoT platform could be chosen.

Rockchip's RK3566 sits right in that price range.

Laying Out the Chip's Specs

Within Rockchip's product line, the RK3566 is positioned as a mid-range AIoT chip, with its main battlegrounds being education tablets, digital signage, dictionary pens, and robot vacuums. This Microduck is a new scenario for it—and this new category has once again put Rockchip in the spotlight. Let's pull out a few key items from its spec sheet and map them against what Microduck needs:

Once you lay out this table, anyone who has worked on RK platforms will recognize the pattern: the CPU runs the main OS, the NPU handles small-model inference, the multimedia block processes video, and the security block does OTA signature verification—this is the classic RK3566 playbook. But a few of Microduck's numbers are quite demanding on the chip, so let's break them down one by one.

The Most Striking Coincidence—16 PWM Channels for 15 Motors

The moment I saw "15 motors" in Microduck's specs, I genuinely smiled. Why? Because on the datasheet of the RK3566 I have here, the number of PWM channels is listed as 16.

The 15 motors hang off the 16 PWM channels, leaving one channel as redundancy or for expansion (say, a future extra wing, or an active cooling fan). This isn't a solution patched together after the fact—it's the chip's capability just perfectly covering the product's needs. Anyone who has done hardware selection knows how satisfying that "one channel to spare" margin feels.

Motor control needs more than just the PWM line:

l 1 PWM channel controls one servo's angle

l Direction control: usually one more GPIO (some servos already build the direction logic in)

l Over-current detection: via SARADC (the RK3566 has 4 channels)

l Position feedback: (advanced) via ADC / I2C

l Multi-device sync: via the PWM sync trigger or SYNC signal

Beyond PWM, the RK3566 also packs 142 GPIOs, 6 I2C, 4 SPI, and 10 UART—peripheral density is its most plain-and-simple advantage. In other words, the first reason Microduck chose the RK3566 is that it is itself a chip "designed for peripheral-intensive products." Motors, sensors, I2S microphone arrays, ToF LiDAR—everything can be hooked on without adding a single I/O expander chip.

What Can a 1 TOPS NPU Actually Do?

Many reposted articles say it "comes with an AI accelerator," which sounds impressive. But most people actually have no idea what the number 1 TOPS really means. Let me first draw a clear line between what it can do and what it can't:

What it CAN do (the era of small models):

l Voice wake-word detection (VAD + keyword, <50 MB model)

l Simple visual classification (is there a person? is there an obstacle?)

l Small policy networks trained by reinforcement learning (Microduck's posture control)

l 1–2 small sensor-fusion networks

What it CANNOT do:

l LLM inference (even a 0.5B model needs at least 4 TOPS)

l Large vision models (something like YOLOv8-large needs 10+ TOPS)

l Complex SLAM / real-time point clouds

Microduck's on-device AI is used where it matters most—it doesn't run any big model, but rather the control policy trained through reinforcement learning. This path is clear:

1. Train the gait policy in a MuJoCo simulation with 4,096 parallel environments

2. Export to ONNX

3. Convert it in one click to an RK3566 NPU RKNN model with RKNN-Toolkit2

4. Run inference on embedded Linux with rknn_runtime, outputting the angle target for all 15 motors every frame

This is the payoff of the toolchain investment Rockchip has made over the past few years. RKNN supports almost all mainstream frameworks—TensorFlow / TFLite / PyTorch / ONNX / Caffe / Keras / Darknet—for "one-click quantization + compilation + deployment."

Perception Interfaces (Camera, LiDAR, IMU, Microphone)

For a robot to "see the world," the RK3566's interface list is basically built for this:

l 1 camera → MIPI-CSI-2 4 lanes, with an 8 MP ISP, for image capture and basic ISP

l 8×8 ToF LiDAR → hooked up via SPI or I2C to run ranging frames

l Dual IMU → each takes one I2C bus (IMU is almost standard on I2C / SPI interfaces)

l Microphone array → the RK3566 generously provides multiple I2S/TDM/PDM digital audio interfaces

Especially for audio, the RK3566 is quite generous with interfaces—a product that needs to "hear sounds around the home" must at least have a PDM microphone-array interface. This Microduck doesn't yet do far-field voice pickup, but the interface is reserved, so adding a mic array later won't require changing the chip.

Interface selection has an easily overlooked trap—whether a chip's BOM includes a certain interface and whether you're willing to use it are two different things. A chip like the RK3566 that has "more than enough" interfaces means the solution provider's hardware design has high fault tolerance—if one path doesn't work, you can switch to another.

Why Not the RK3588?

This is what many readers would ask first—since we're building a robot with on-device AI, why not use the RK3588? Isn't a 6 TOPS NPU more appealing? I have to explain this from a hardware-selection angle:

First, the price can't hold up. The total BOM of an RK3566-based device can often be pressed down to double-digit dollars. The RK3588 would cost noticeably more at an equivalent BOM. With the whole Microduck at $399—and it still has to pack in 15 motors + LiDAR + dual IMU + a battery—the budget left for the SoC is very limited.

Second, the power budget can't hold up. The RK3588 uses an 8 nm process, but at full load it still consumes far more power than the 22 nm RK3566. For an 800 g duckling, battery capacity has a hard limit—low power is worth more than peak compute.

Third, it's actually overkill. That reinforcement-learning control policy is about 1 MB after quantization, and on 1 TOPS it can still run several hundred frames per second—performance is massively excessive. Putting a 6 TOPS NPU in a Microduck is like dropping a V12 engine into an electric scooter—it'll run, but it's a waste.

Fourth, package size. The RK3566 uses a 15.5 mm × 14.4 mm FCCSP565L package (0.4 mm + 0.6 mm ball pitch), a size smaller than the bulky RK3588 package. Inside a 25 cm tall robot, every millimeter is expensive.

I've been working with the RK3168, RK3188, and RK3128 since 2015, and the biggest lesson is—don't chase the top spec. Top-tier chips' compute is genuinely great for big models and desktop vision, but on small edge devices, mature "good enough" beats "excessive" every time.

It's Not Just a Chip—It's an End-to-End Platform

Hardware done, let's look at the other half of why Microduck went viral—the software ecosystem. A lot of media only saw the result "Rockchip hit limit-up" and missed the quiet, long-term work Rockchip has been doing on its developer ecosystem. Let me show you a few numbers:

l Rockchip's official GitHub account: 39 public repositories, about 16,800 stars

l Three AI projects—rknn-toolkit2 / rknn_model_zoo / rknn-llm—grew their stars from 1,821 to 7,687 in two years, a 322% increase

l On Hugging Face, models found by searching the keyword RK3588: 386 in total, of which 96.9% were created in the past two years

Rockchip's toolchain directly consumes the Hugging Face ecosystem. A developer pulls a pretrained model from HF → exports to ONNX → quantizes with RKNN-Toolkit2 → deploys to the RK3566 NPU in one click. This pipeline needs no low-level driver writing, and its barrier to entry is far lower than calling the vendor SDK directly.

The Microduck team announced "fully open-source SDK + the entire simulation training stack opened to the public." But the open-source hardware-platform ecosystem behind it is something Rockchip grew slowly over several years.

Some Thoughts

Finally, something a bit softer. I've been in embedded engineering for a decade, and I've seen every generation of the RK series from the RK3168 to the RK3588; I've seen many chips work quietly their whole lives without ever making the news. This time the RK3566 went viral via Microduck, and it gave me some unexpected feelings—

1. Domestic AIoT SoCs have moved past the "shop around for the cheapest" stage and entered "value-based pricing." Microduck isn't cheap at $399; Rockchip's RK3566 doesn't win on low price either. This wave trades platform capability for a premium—no longer the old race to the bottom on cost.

2. The real battleground of edge AI is not cloud LLMs, but small devices like this with 1 W power and around $100 cost. The iPhone moment for big models hasn't come, but the robot moment for small models is happening now. Microduck's reinforcement-learning control policy is only 1–2 MB, yet it gives this duck a "sense of life"—and that's exactly what users are paying for.

3. The greatest happiness for an embedded engineer is seeing a chip you know well become needed by the world. The first time I played with the RK3168 was in 2015, when domestic SoCs were still questioned with all kinds of "domestic substitution" doubts. Ten years on, domestic chips can keep up with the product cadence of a global developer community's hit product—that sense of achievement isn't something written on a paycheck.

Appendix: What Does This Duck Actually Look Like?

You can't tell this duck's "character" from a spec sheet alone. I pulled a few unboxing photos from Pollen Robotics' official press kit to show you—

l Close-up: white shell + orange beak + monocular camera on the head

l Workbench in the morning light

l The roller-skate version—yes, the $399 duck can also ride a skateboard

Article shared from the Internet

ONE MORE THING

Thinkcore Team is an embedded board solution provider, developer, and motherboard supplier, specializing in the design, manufacturing, sales, and technical sharing of embedded motherboards. Our chipset portfolio includes Rockchip, Allwinner, and others. We welcome inquiries and collaboration!

mail:  developer@thinkcore.cn