Not a parts catalogue you browse and abandon — an ordered path from "I want to build a robot" to a machine that closes a control loop, senses itself, and does something repeatably. Eight steps, the parts that fit each one, and the failure modes nobody warns you about.
The order matters more than the parts. Each step below assumes the previous one produced something you can measure.
Not a humanoid. A robot with one actuator, one sensor and one closed loop, that fits on a desk and runs on a single power supply. Your first build exists to teach you the integration loop: CAD → print → wire → power → firmware → feedback → debug. Choose a pan-tilt camera, a line follower, or a 2-DOF arm. Finish it.
This is the step nearly everyone skips, and it is the reason their robot is un-debuggable. An encoder and an IMU cost less than a single servo and they turn "it moves wrong" into "joint 2 is 12° behind command at 400 ms". Without feedback you cannot tell a mechanical fault from a firmware bug, and you will guess for weeks.
Every motor and sensor on a table, wired to the controller, driving one joint. Write the loop to command a position and log commanded versus actual. If it cannot hold position on a bench with nothing attached, it will not hold position on the robot with gravity attached. Do not skip to mounting because the bench feels unrewarding — the bench is where the project is won.
Write down bus voltage, worst-case current per actuator, total peak current, and required runtime. Then pick chemistry and capacity with 1.5× margin, and put a fuse plus a real switch upstream of the pack. Most "the controller randomly resets" problems are a buck converter browning out because the pack sagged under peak load. Draw the tree as a diagram; do not improvise it with a breadboard.
Work backwards from the end effector: load mass × link length gives required torque at each joint moving inward. Compare that to continuous motor torque and pick a ratio. Then decide what provides it — a gearbox, a belt, or a printed cycloidal — knowing each stage costs you backlash and efficiency. A 5:1 belt and a good encoder beats a 27:1 gearbox with 3° of slop for arm accuracy.
One microcontroller owns the control loop at a fixed rate; one Linux or GPU board owns perception, planning and the network. They exchange a small, defined message set. Never let Wi-Fi, ROS 2 or a Python process sit inside a 1 kHz loop — that is how robots twitch. This boundary is the single most important architectural decision in the whole build, and it is cheap to get right on day one and expensive to retrofit.
Printed or extrusion for v1, always. You will change the geometry at least twice and both materials make that cheap. Stiffness is a v2 problem, after the kinematics and the control loop work. When you do upgrade, remember that bolted extrusion joints creep under vibration and that carbon fails catastrophically with no visible warning — match material to the actual load path.
Log commanded vs actual position per joint, motor current, loop timing jitter and pack voltage. Every improvement after this point should be justified by a number in that log, not by intuition. The three failures that quietly eat months are backlash compounding across a serial chain, power sag under peak current, and a servo asked to hold stall torque continuously until it cooks its own windings. All three are invisible in CAD and obvious in a log.
Pick the row that matches what you can spend. All three use the same roadmap — only the parts scale.
| Starting point | First build | Actuation | Feedback | Parts cost | Time to first loop |
|---|---|---|---|---|---|
1 Desk learnerabsolute beginner |
Pan-tilt camera or line follower | 2× micro servo | AS5600 + no IMU yet | $35 – $60 | one weekend |
2 Arm builderknows some code |
SO-101 5-DOF arm + teleop | 6× Feetech STS3215 | Serial encoders in servos | $110 – $250 | 2 – 3 weeks |
3 Mobile roboticswants ROS + navigation |
Diff-drive AMR with Nav2 | 2× DC gearmotor + encoders | RPLIDAR A1 + BNO085 | $450 – $900 | 4 – 6 weeks |
A $110 arm is $110 in servos and filament. Add the USB serial adapter, a power supply, the tools you do not own, a camera, and the parts you reprint after a mistake, and first-working-demo is realistically $250. Every project on this site has that gap between BOM and reality.
Seven subsystems. Everything in the catalogue is one of these, and every build is a negotiation between them.
The muscles.
Gearboxes, belts, lead screws and strain-wave reducers.
Proprioception and exteroception.
Two tiers, always.
Cells, BMS, regulators, e-stop.
Extrusion, carbon tube, machined plate, printed polymer.
Where the robot meets the task.
Part count is not cost. This is the spread of the catalogue by subsystem — count on the left, the cheapest and the dearest entry in each on the right.
The cheapest one is $110 and the most expensive is $150k, but the architecture is the same. Read the cheap ones for mechanics, the expensive ones for what a mature stack looks like.
The Robot Studio / Hugging Face · 2025
The $100 arm that changed who gets to do robot learning. A five-bar printed arm driven by six STS3215 serial servos, trained by teleoperation — you move it, it copies you, then it repeats. This is the hardware that made imitation learning a hobby.
UC Berkeley (open) · 2025
A full humanoid designed explicitly to be built by anyone with a 3D printer and a credit card. Printed cycloidal reducers, off-the-shelf BLDCs, and every CAD file published.
Juan Miguel Jimeno (open) · 2022
The most-documented open mobile robot base: a diff-drive chassis with an ESP32/Teensy micro-ROS bridge, a LiDAR, and a complete ROS 2 Nav2 stack that works out of the box.
Unitree · 2023
The consumer quadruped that made legged robots normal. Its real achievement is the price: twelve QDD joint motors, a LiDAR, and a full locomotion controller for less than a used car.
Clearpath / Open Robotics · 2023
The reference ROS 2 teaching robot. Pi 4 + OAK-D + RPLIDAR A1 on a Create 3 base, with every tutorial in the ROS 2 ecosystem written against it.
Franka Emika · 2017
The arm that defined torque-controlled manipulation research. Seven DOF, joint torque sensors in every joint, and an interface that lets you command torques at 1 kHz — which is why it is in so many papers.
If you buy nothing else, buy these to get through steps 2 and 3 — feedback, a brain, and a way to see what is happening.
A 12-bit absolute magnetic encoder on a 3 mm chip, for the price of a coffee. Stick a diametric magnet on the shaft and you have joint feedback.
Does the sensor fusion for you. Onboard ARM core outputs a quaternion, gravity vector and linear acceleration — no Kalman filter to write, no magnetometer calibration nightmare.
Dual-core 240 MHz with Wi-Fi and BLE built in. Runs the control loop, drives the servos, and serves a web UI over the air for five dollars.
Measures bus voltage and current across a shunt and reports both over I²C. Ten lines of code buys you the ability to see a joint stalling before it burns.
5 V at 5 A from up to 38 V in, with a genuinely honest current rating and good thermal behaviour. The regulator you buy when you are tired of cheap modules browning out your Pi.
The workhorse of the low-cost arm movement. 4096-step magnetic encoder, TTL half-duplex daisy chain, aluminium gears on the 30 kg·cm variant. This is the servo inside SO-100 / SO-101 and most hobby legged bots.
Backlash, which compounds across a serial chain until the end effector is centimetres off. Power sag, where the pack cannot supply peak current and the controller browns out mid-motion. Thermal, where a servo rated for stall torque is asked to hold it continuously and cooks its windings. All three are invisible in CAD and obvious in a log.
Model the mechanism in MuJoCo or Isaac Lab with realistic masses and torque limits before committing to an actuator. A simulation that says your knee needs 40 N·m peak and your chosen motor delivers 14 is the cheapest engineering you will ever do — and it takes an afternoon.