Degrees of Freedom (DOF)

DOF is the joint count. A 4-servo biped can waddle; 16–18 DOF gives you real arms and a hip that can rotate; 20–24 DOF adds grippers and a waist. More DOF means smoother gaits and more interesting projects, but also more calibration, more current and more to break.

Servo Type & Gears

Plain PWM servos are cheap but blind. Serial bus servos daisy-chain on a few wires and report position, load, voltage and temperature back, which is what makes closed-loop gait control possible. Metal gears are non-negotiable on the legs: a plastic hip gear will strip the first time the robot falls.

Controller Platform

micro:bit and Arduino kits are great for learning motion sequences. An ESP32 adds Wi-Fi and enough power for a camera module. A Raspberry Pi 5 unlocks OpenCV vision, Python, ROS and voice models. Match the brain to what you actually want to learn, not to the spec sheet.

Gait & Inverse Kinematics

The difference between a toy and a platform is whether the software solves inverse kinematics: you give it a foot position, it computes the joint angles. Kits with IK and omnidirectional gait planning let you write real behaviours; kits with only recorded action files limit you to playback.

Docs, SDK & Spare Parts

A humanoid is worthless without documentation. Look for a published SDK or GitHub repo, step-by-step assembly guides, and, critically, replacement servos and brackets you can actually buy. Closed platforms with no secondary development support will stop being fun the moment you outgrow the app.

The Unitree G1 is what happens when a humanoid stops being a kit and becomes a machine. It stands 1,320 mm tall, weighs about 35 kg, and carries 23 joint degrees of freedom: six per leg, five per arm and one at the waist, driven by low-inertia PMSM motors with crossed-roller output bearings and up to 90 N·m at the knee. An 8-core CPU, depth camera and 3D LiDAR handle perception, Wi-Fi 6 and Bluetooth 5.2 handle connectivity, and a 9,000 mAh quick-swap battery gives roughly two hours. Be clear-eyed about two things, though: this sits far above every other pick here on price, and the version sold on Amazon explicitly does not support secondary development. If you need SDK or API access you want the G1 EDU instead.

If your goal is research or a university course rather than a weekend build, the AiNex is the pick. It is built on ROS, runs on a Raspberry Pi 5, and packs 24 intelligent serial bus servos, enough for movable mechanical hands and two extra hip joints that let the legs rotate. That extra hip DOF is what enables genuine omnidirectional walking rather than a fixed forward shuffle, and Hiwonder ships an inverse kinematics layer plus gait planning so you can command poses and trajectories instead of hand-tuning angles. Control comes from the WonderROS app, a wireless gamepad or PC software.

The Tonybot is the sweet spot for anyone who lives in the Arduino world. It runs on an ESP32 with 17 intelligent servos, walks, dances, kicks a ball and, the party trick, stands itself back up after a fall. An add-on ESP32-S3 vision module brings offline face recognition, target tracking and visual line following, and a voice module handles spoken commands; the higher bundles add multimodal AI models for conversation and scene understanding. You program it in Arduino C++, Scratch-style blocks or Python, and an ultrasonic sensor plus expansion pins keep it hackable.

The XiaoR Geek XR-MRT is the fun one. It is a 27 cm, 18 DOF humanoid with metal-gear servos distributed as one in the head, one in the chest, six across the arms and ten in the legs, a genuinely capable joint layout for boxing, football and dance routines. What sets it apart is the swappable armour: the exterior panels and servo colours are meant to be customised, so it is the kit to buy if you want a robot that looks like yours. It supports secondary development through a documented interface, and you program it graphically from the phone app or Windows PC software.

Built around the BBC micro:bit, Robot PU is the humanoid designed for teaching rather than tinkering. It self-balances, walks, dances, plays soccer and runs an autopilot mode that avoids obstacles and even interacts with other PU robots, which is perfect for a class set. The software path is the real selling point: MakeCode blocks for beginners, then JavaScript and Python as students progress, with the firmware open-sourced on GitHub and 30+ free online lessons covering programming, 3D design, printing and AI. Expansion kits add construction parts, extra pins and an AI camera.

ELEGOO doesn’t build a full humanoid, and we’d rather say so than pretend otherwise, but the Penguin Bot V2.0 is the best cheap way to learn how a walking biped actually works. It is an Arduino-based kit with servo-driven legs, obstacle avoidance, auto-follow, Bluetooth app control, a music player and dance routines, and it ships with a proper illustrated manual, spare screws and tools. Crucially it is fully reprogrammable in the Arduino IDE, so once you’ve built it you can rewrite the gait, add an ultrasonic behaviour or bolt on your own sensors. Think of it as the SG90-class entry point to bipedal robotics.

Want the full experience?

The Hiwonder TonyPi Pro gives you 20 DOF, a Raspberry Pi 5, camera vision and voice control in one documented package.

Want AI vision for less?

The TonyPi Standard runs the same Python and OpenCV stack on a 4GB Pi 5 for noticeably less money.

Doing ROS or a thesis?

The Hiwonder AiNex is a native ROS platform with 24 bus servos, inverse kinematics and omnidirectional gait.

Live in the Arduino world?

The ESP32-based Tonybot gives you 17 DOF and optional offline vision without touching Linux.

Teaching a class?

Robot PU runs on a micro:bit with MakeCode blocks, an open-source repo and 30+ free lessons, ideal for a class set.

Just want to try walking?

Start with the ELEGOO Penguin Bot V2.0, a cheap Arduino biped you can rewrite from scratch.

How many degrees of freedom does a humanoid robot kit really need?

For a robot that only has to walk forward and turn, four to six leg servos is enough, which is the Penguin Bot class. To get expressive arms and a hip that rotates you want 16 to 18 DOF, which is where the Tonybot and XR-MRT sit. Above that, the extra joints buy you grippers, a waist and true omnidirectional gait: the TonyPi Pro’s 20 DOF and the AiNex’s 24 servos exist so the robot can pick things up and side-step, not just walk. More DOF also means more calibration, more current draw and more that can strip in a fall, so pick the lowest count that covers the projects you actually want to build.

Can I actually program the Unitree G1 I buy on Amazon?

Not in the developer sense. The G1 listing sold on Amazon states plainly that secondary development is not supported. You get app and remote control, voice interaction, a library of official motions and OTA updates, but no SDK or API access, and some complex maneuvers are locked for safety. If you want to write your own controllers, run reinforcement-learning policies or integrate the robot into a research stack, you need the G1 EDU version, which Unitree sells separately. Buying the consumer G1 and expecting to code on it is the single most expensive mistake you can make in this category.

Why do these kits use serial bus servos instead of normal PWM servos?

Two reasons: wiring and feedback. An 18-joint robot wired with standard PWM servos needs 18 signal lines plus a driver board, and none of those servos can tell you where they actually are. Serial bus servos share one TTL line, each with its own ID, so the whole leg chain runs on three wires, and each servo reports back position, load, voltage and temperature from an internal encoder. That feedback is what lets the controller detect a stalled joint, hold a pose under load and run closed-loop gait correction. It’s the same idea behind smart servos in open-source robot arms, and it’s why bus-servo kits cost more.

Do I need a Raspberry Pi, or is an Arduino/ESP32 humanoid enough?

It depends entirely on whether you want vision. An ESP32 or Arduino can drive 17 servos, run a walk cycle, read an ultrasonic sensor and take Bluetooth commands perfectly well, and the toolchain is far simpler. But the moment you want OpenCV, MediaPipe pose detection, ROS nodes or a speech model, you need a Linux board. That’s why the TonyPi and AiNex ship with a Raspberry Pi 5. A good rule: choose ESP32 if the project is about motion and control, choose Raspberry Pi if the project is about perception.

How hard are humanoid kits to assemble, and will it fall over?

Assume a full day for an 18–24 DOF kit, most of it spent centring servos before they go into brackets. Get that wrong and the joint will hit its end-stop mid-stride. And yes, it will fall over: bipedal walking is unstable by nature, floor surface changes the gait, and a low battery makes the servos sag. Build on carpet or a mat, keep the battery charged, buy one or two spare servos and expect to re-tune the gait after any hard landing. That whole loop is the point of the hobby: you’re learning the same balance and calibration problems that full-size humanoids face.