GrowBot
What's the simplest possible AI robot that learns everything from scratch?
GrowBot is a palm-sized, low-cost bipedal robot built on a Raspberry Pi Zero 2 W that uses tiny neural networks for walking and an AI language model to decide what to do next, a minimum viable physical AI you can 3D-print and build yourself.
GrowBot is a small, low-cost bipedal robot built around a Raspberry Pi that the creator describes as a minimum viable physical AI. The project starts from a question: if you strip a humanoid robot down to the absolute bare minimum in parts and cost, what do you get? The answer is a palm-sized two-legged robot with a camera, a microphone, a motion sensor, a small speaker, an LED ring, and two servo motors for legs, all powered by a single small battery.
The hardware runs entirely on a Raspberry Pi Zero 2 W, a credit-card-sized computer that costs around fifteen dollars. There are no additional computers. The legs use a type of servo motor that reports its own position and load back to the Pi, which lets the software know how the legs are actually moving rather than just assuming. A wiring diagram and a full parts list with prices are included in the repository.
The software design has two layers. The lower layer consists of small learned motion policies: compact neural networks trained offline that output servo position targets directly, rather than relying on a hand-written walking pattern. These are deployed onto the Pi after training and continue learning from each run. The upper layer is an agent loop that reads sensor data from the camera, microphone, motion sensor, and servos, sends that picture to an AI language model via an API, and lets the model decide on a goal and an action. The robot tries the action, feeds the result back, and the loop continues.
The README is candid that this is a very early snapshot, labeled V0. The power rail design is acknowledged as a rough shortcut, the full training code is not yet included, and the repository is described as a reference for a build video rather than a step-by-step beginner guide. A more stable V1 is planned for late 2026 with proper build instructions, a custom circuit board, and a calibrated simulation model.
The 3D-printable body files and a MuJoCo physics simulation file for the body are included. The project is released under a Creative Commons license that allows non-commercial use with attribution.
Where it fits
- Build your own tiny walking robot at home using 3D-printed parts and a cheap Raspberry Pi.
- Experiment with AI controlling a real physical robot, the language model picks goals and the robot acts them out.
- Learn how small neural networks can teach a robot to walk without hand-coded movement rules.
- Use the included physics simulation to test and train robot movements before building the real hardware.