← Build

Build

The CyberBrick Rover

One printed body, one board of your choosing, one policy, proven end to end. Drive the twin now, then take the same code to real hardware. This is the smallest complete loop in Physical AI.

from ~$45cost
brain-agnosticboard
FDM · any printerbody
first buildlevel
The sim twin

Run a real policy, right here.

The four-part build starts with this: the design in real MuJoCo physics. What drives it here is what drives it on your bench.

MUJOCO · REAL PHYSICS

Drive the twin before you print it

This is the CyberBrick Rover in real physics. Write a policy, run it, and it drives to the gold pad, the same control(obs) you'll stream to the board.

You read

obs each tick: goal_dist, goal_bearing, range, heading, t

You change

the rules inside control(obs): the policy itself

Fixed

the rover, the arena, and the MuJoCo physics

What you build

Four parts, one loop.

Body, brain, driver, policy. The design forces no single board, only that whatever you pick can take the commands and run the policy.

01 · Body

A printed two-motor chassis

A skid-steer rover you print on any FDM machine: chassis, two geared-motor mounts, a caster, and a snap-in board tray. Bambu / MakerWorld-anchored, no exotic hardware.

02 · Brain

Whatever fits your budget

Skid-steer needs two motor channels and a way to take commands, so almost any board here works. Pick from the options below; the design locks to none of them.

03 · Driver

Browser → board, over one cable

The policy you write on this page streams left/right track commands to the board over WebSerial (USB-C) or BLE, the same driver registry the Institute's sim-to-real bridge uses. The board is the hands; your laptop is the brain.

04 · Policy

The code in the sim is the code on the robot

control(obs) → two motor commands. Train it against the twin above, then point the same function at the driver. Perception can run on-device or offload to your webcam.

Pick a brain

Ways to power the same body.

Cheapest to heaviest. Every one drives the identical printed body and runs the same policy.

Bambu CyberBrick core (ESP32-C3)

~$45 kit

The cheapest path. Two motor ports on the receiver shield; runs a small MicroPython listener while your laptop runs the policy and streams commands. Bodies come from the same MakerWorld community.

Espressif ESP32-S3

~$8

A bare, fully-open board when you'd rather flash your own firmware. Add a motor driver; run tiny vision on-device or offload it.

Raspberry Pi 5 (+ AI Camera / Hailo)

$80+

The heavy option: full on-board perception and a real policy running on the rover itself, no tether. Overkill for a first build, but the same body carries it.

Full specs and roles for every board: Digital Brains →

In the wild

Open robots like this one.

The Maker Atlas maps real, open Physical-AI robots to each of our builds. These are kin to this one, open a profile to trace its boards, sensors, and source.

In the wild

LeRobot LeKiwi

A holonomic mobile base + SO-101 arm = a ~$1k open mobile manipulator on a Raspberry Pi 5, dual cameras. LeRobot.

In the wild

OpenBot

A ~$50 printed rover that uses an Android phone as the brain (camera + IMU) + an Arduino MCU; TFLite person-following. MIT, ~3k★.

In the wild

DonkeyCar

The open self-driving RC car, Raspberry Pi or Jetson + a wide-angle camera, behavioral-cloning autopilot. MIT; the DIY-Robocars standard.

In the wild

NVIDIA JetBot

The reference open AI robot on Jetson, CSI camera, PyTorch notebooks (collision-avoidance, road/object-following). MIT, fully FDM.

In the wild

Duckietown

An autonomy-education mobile robot + miniature city, Jetson/Pi + a fisheye camera, ROS lane-following. Open, used in MIT/ETH courses.

In the wild

SunFounder PiCar-X

A Raspberry Pi vision/AI rover, a pan-tilt Pi camera doing color, face, and line tracking. Open, popular in education.

In the wild

NASA JPL Open Source Rover

NASA JPL's open Perseverance-style 6-wheel rocker-bogie rover, Raspberry Pi, add-your-own camera/IMU. Apache-2.0.

In the wild

NVIDIA JetRacer

NVIDIA's open Jetson-Nano autonomous racing car, a wide-angle CSI camera + PyTorch. MIT, the Jetson-education line.

In the wild

LEGO Education SPIKE Prime

The current LEGO robotics platform: a Powered Up hub, encodered angular motors, and colour/force/distance sensors. Excellent body and actuation; no camera and no on-hub ML.

In the wild

micro:Maqueen + HuskyLens

A micro:bit rover chassis plus HuskyLens, a snap-together robot that sees and tracks with an on-device neural camera today. The closest kids' kit to real physical AI.

In the wild

Cubelets

Magnetic robot blocks: every face carries ground, power, and data, and behaviour emerges from how you snap Sense, Think, and Act blocks. The embodied loop in your hands, no code, a hard ceiling at no camera and no learned policy.

See how they all connect: the Maker Atlas →

Where this build stands

Live now: the twin. In progress: the physical unit.

The sim twin on this page is real and live today. The printable body and the driver are specced against the sim-to-real registry; the first physical unit is being assembled, and the STL pack and step-by-step land here the moment it's verified on real hardware, not before.

3D printing guide →The design method →

Make it

Design its body, program its brain.

Take this build into the Studio: model and export its printed parts on the CAD kernel, and write the MicroPython that runs on its board. Both open as full-screen workspaces.

Design the body →Program the brain →