Most robot dogs on a desk are toys, and most serious quadrupeds cost as much as a car. Hiwonder’s PuppyPi sits deliberately in between: a 720-gram, aluminium-framed quadruped built around a Raspberry Pi 5 that you can program in Python, run on ROS 2, and now talk to through ChatGPT-class large models. Say “find the red ball and kick it,” and the robot is meant to hear you, look around, work out what you mean, and do it.
That chain, from words to perception to motion, is what researchers call embodied intelligence. PuppyPi packages a small, hackable version of it for classrooms, makers and anyone who wants to learn how modern robots connect language models to the physical world.
What’s inside the PuppyPi
According to Hiwonder’s specifications, PuppyPi measures 226 × 149 × 190 mm, weighs 720 g and runs on a 7.4 V 2,200 mAh LiPo battery. Its legs use a linkage design driven by eight HPS-0618SG coreless servos with stainless steel gears, giving it 8 degrees of freedom. That is two per leg, enough for walk, amble and trot gaits plus stair climbing (Hiwonder).
A wide-angle camera in the head supplies first-person video (the spec sheet lists 480p), and an IMU lets the robot sense and correct its posture. Yanko Design, which covered the robot in 2025, put the servos at 8 kg·cm of torque each and the runtime at around 60 minutes per charge (Yanko Design).
The brain is a Raspberry Pi 5, offered with 2 GB to 16 GB of memory, or an older Raspberry Pi 4B. The choice matters: Hiwonder states that the Pi 5 versions run both ROS 1 (Ubuntu 20.04 with Noetic) and ROS 2 (Ubuntu 22.04 with Humble), while the Pi 4B version supports ROS 1 only (Hiwonder). An 8 GB Pi 5 gives comfortable headroom for running vision, ROS 2 nodes and model calls at the same time.
How the ChatGPT integration actually works
The “large AI model” label is easy to misread. The language models do not run on the robot. Hiwonder’s ROS 2 course shows a cloud pipeline built from familiar pieces (Hiwonder docs):
- Wake word: the WonderEcho Pro voice box, a USB microphone and speaker with noise suppression, listens locally for “Hello HiWonder.”
- Speech to text: your command is sent to OpenAI’s speech recognition (the examples use whisper-1 and gpt-4o-transcribe).
- Reasoning: the text goes to an OpenAI language model that turns it into structured actions the robot’s ROS nodes can execute, for example “take two steps forward, turn left and take one step back.”
- Seeing: camera frames go to a vision language model through OpenRouter, so the robot can describe a scene or locate an object you named.
- Talking back: replies are spoken through OpenAI text-to-speech.
In practice that means you need your own OpenAI and OpenRouter API keys with credit, plus a working internet connection. Hiwonder’s docs walk through creating both keys. The upside is that you can swap in newer models as they appear, since the code is open.
What it can do out of the box
Hiwonder lists a long menu of demos (Hiwonder):
- Classic vision with OpenCV and MediaPipe: colour recognition, target tracking with PID control, AprilTag-style tag recognition, face detection, line following, autonomous hurdling and stair climbing.
- Language-driven tasks: voice control, scene understanding (describing what the camera sees), ball tracking and kicking on command, patrolling, posture reports from the IMU, and temperature and humidity readouts when the sensor kit is fitted.
- With add-ons: a TOF LiDAR for 360° scanning, SLAM mapping, navigation and dynamic obstacle avoidance, and a small robotic arm for grasping and object transport. Hiwonder told one customer the arm’s payload is about 30 g.
You can also skip code entirely: Android and iOS apps, PC software and a wireless gamepad all drive the robot, and it ships assembled.
Standard, Advanced, Pro or Ultimate: pick carefully
This is the detail most product listings bury. According to Hiwonder’s own packing lists, the Standard kit includes the robot with camera, charger, card reader, balls and tags, but not the WonderEcho Pro voice box (Hiwonder docs). The Advanced kit adds the WonderEcho Pro. The Pro kit adds a wireless controller, sensors (ultrasonic, temperature and humidity, touch, dot-matrix display) and the TOF LiDAR. The Pro Ultimate kit adds the robotic arm.
So a “Standard + Raspberry Pi 5 8GB” PuppyPi gives you the full ROS 2 and computer vision stack, but voice-driven ChatGPT interaction needs the voice box, either by choosing the Advanced kit or adding the module later. Hiwonder lists PuppyPi from $589.99, with the price changing by kit and controller (Hiwonder).
Who it’s for, and the caveats
PuppyPi is an education and research platform, not a robot pet. Hiwonder pitches it at high school and university students, and its open-source ROS 2 repository on GitHub includes packages for control, SLAM, navigation, the arm and large-model integration (GitHub). Gazebo simulation support lets you test gaits and algorithms before touching hardware.
Keep expectations grounded:
- Linux comfort helps. The best features live in ROS 2, Python and the terminal.
- Cloud costs and connectivity. Voice and scene understanding depend on paid API calls and a network.
- Local network only. Hiwonder has said the robot can’t be controlled when you’re away from home.
- Servos wear. The manual calls them consumables and advises letting them cool after long sessions.
Key takeaways
- PuppyPi is a compact, aluminium Raspberry Pi 5 quadruped with 8 coreless servos, a camera and an IMU, running ROS 1 and ROS 2.
- Its “ChatGPT” features are a cloud pipeline: local wake word, then OpenAI speech recognition, language model and speech, plus a vision language model via OpenRouter.
- The Standard kit lacks the voice box. Choose Advanced or higher if voice interaction is the point.
- The Raspberry Pi 5 8 GB option is the sensible middle ground for ROS 2 plus vision workloads.
Conclusion
PuppyPi won’t fetch your slippers, but it does something more useful for learners: it makes the full embodied-AI loop of hearing, seeing, reasoning and moving visible and editable on a robot that fits on a desk. For students, educators and makers who want to see how language models become motion, it’s one of the most complete small platforms available.