B grade Authoritative secondary, traced to origin

Build a servo quadruped: the Spot Micro route to an MVP

The legged route most people actually finish. The MVP is not smooth walking; it is stable standing and a straight-line static-gait walk without falling over.

Type: Quadruped, 12 DOF (position control, no force control)
Budget: About USD 200–350
Servos: MG996R / DS3218 × 12 (metal gear, PWM)
Electronics: PCA9685 + Raspberry Pi + MPU6050
Software: Python gait + inverse kinematics (SpotMicroAI route)
MVP check: Static-gait walk over 2 m without falling

Servo quadruped: how the body goes togetherTap a number — or the legend below — to jump to its entry

What counts as moving: MVP = a static gait (every phase statically stable) over 2 metres of straight line, plus a squat-and-rise. Checks: the body sits level at stance — verified with IMU readings, not by eye; no leg collapse or foot slip within a gait cycle; servos not hot to the touch after ten cycles. State one thing plainly first: this machine has no force control — it is position control only. That means the ceiling on 'how well it walks' was set the day you finished assembly; you tune parameters within it, never beyond it. Making peace with that makes tuning far more pleasant.
Parts and sourcing: Twelve servos (MG996R or DS3218, metal gear, USD 5–8 each; not SG90s for legs — plastic gears under leg load are a scheduled failure, not a risk), a PCA9685 16-channel PWM board (about USD 2), a Raspberry Pi (3 or newer), an IMU (MPU6050), and a 5–6 V high-current supply (power is the quadruped's trap number one: one stalled servo can pull 2.5 A, and the peak of twelve acting together has to be budgeted. Undersized power shows up as 'gets softer as it walks', not 'does not move' — a symptom that has fooled nearly every beginner). Printed parts, about 500 g. USD 200–350 all in.
Assembly order: Print → build one leg on the bench first (finishing all twelve before discovering a homing problem is the most expensive beginner move) → home each servo with a fixture, not by eye → four legs + body → wire the PCA9685 (I2C address; servo power strictly separate from the Pi) → mount the IMU at the body's geometric centre. Everything hinges on servo homing: the finished legs must show the standard stance through mechanical alignment at assembly — software can compensate for error, it cannot compensate for a mistake.
Software stack: You do not need ROS 2 to start: Python and the community SpotMicroAI codebase (gait interpolation + inverse kinematics). Two roads after: port it to ROS 2 (done by others; navigation attaches, though without force control the payoff is limited); or move to bus servos or brushless joints in a self-designed frame — that is the road where quadrupeds are actually learned. You must understand the IK: copying it is fine, but derive the hip-knee-ankle geometry on paper once. Until then, gait parameters are astrology. This is the line between copying and being able to modify.
Bring-up to first motion: Order: one leg swinging on the bench → four legs standing → level via IMU readings → static gait forward → fall over → adjust → repeat. Three classic traps: ① homing installed wrong — a leg toes in or out at stance; the highest-rework item; ② supply sag — 'gets softer as it walks', mid-walk resets, hot servos; ③ centre of mass too high — tips backward within one cycle; put the battery low and forward. Falling is part of the process: build this machine as something that will fall — no sharp corners on printed parts, and keep a buffer zone off the desk edge.
Where to go next: After MVP: swap in bus servos (Feetech STS reports position and load) — this is the watershed between 'toy quadruped' and 'research quadruped'; force feedback is what dynamic gaits are made of. Or port to ROS 2 (architecturally educational, limited payoff for this machine). Or jump to an open force-controlled quadruped (the MIT Mini Cheetah reproduction route; budget jumps to several thousand). Not recommended: continuing to polish gait parameters on the PWM hardware. The ceiling is in the servos, and it does not move.
What we checked: We verified: the component set (12 servos + PCA9685 + Raspberry Pi + MPU6050) against the SpotMicro-family BOMs; that servo homing and supply sag are the two most frequent problems in community build logs; and the MG996R metal-gear variant's existence and price tier. We did not build this one — fall frequency, the exact voltage threshold behind 'gets softer as it walks', and the step timings come from community reports. The grade stays accordingly.

This site's call

The only machine under USD 300 that gives you real legged-robot intuition. The four things it teaches — IK, homing, centre of mass, power — transfer unchanged to quadrupeds ten times the price. But keep its identity straight: it is a teaching device, not a platform. PWM servos have no force feedback, and position-control gaits are the whole of its world. Take the first lesson, then move up.

'Gets softer as it walks': one symptom, three layers

Layer	What you see	How to check
Power	Overall motion shrinks over cycles, then a reset	Meter the servo rail mid-motion; below 5 V is your answer
Servo thermal protection	Concentrated in a few servos, usually loaded rear legs	Touch them; the hot one is compensating for another
Structure	Parts crack along layers; holes elongate; drift worsens	Inspect layer orientation and hole walls

Why levelling is not done by eye

Human vision is insensitive to 2–3 degrees, but gait is not: 3 degrees of body tilt redistributes leg loading enough to become foot slip within ten steps. The IMU is the only honest referee on the machine.

Sources

SpotMicro-family open repositories (BOM, assembly notes, fork differences)
Community build logs (homing and power sag symptom records)

Objects in this entry

Last checked 2026-09-28