Guide
How robots work: a beginner's guide
The short answer: every robot, from a 3D-printed arm on a desk to a walking humanoid, runs the same loop: sense, decide, act. For a long time the "decide" step mostly belonged to a person, either holding a controller or writing out every move in advance. Most of what's new in robotics is moving that decision into the robot itself, with backdrivable electric motors, better sensors, and AI that learns from simulation and data.
Part 1: Why I'm learning robots in public
Picture a robot about the size of a small dog. It weighs around 20 pounds. It crouches, pushes off, flips backward through the air, and lands on its feet. That's MIT's Mini Cheetah, and in 2019 it became the first four-legged robot to do a backflip (MIT).
When I watched that clip, the question I couldn't shake wasn't "how did they do that?" It was: how does it know where the ground is?
Robots are starting to make their own calls: where to step, what to grab, how to get back up after a fall. They're showing up in warehouses, labs, and on sidewalks.
I want to understand that while it's happening, and I mean how it actually works, not the launch-video version.
I'm not an engineer. I don't have a robotics degree. So I'm teaching myself, and I'm writing down what I learn as I go. That's Servospace.
Why do it in public? Because it keeps me honest. When I write something you're going to read, I have to check it. I have to link where I got it. I have to admit when I don't get it yet. And when I get something wrong, someone will tell me, and I'll fix it where you can see it.
Why should you care? Most robotics writing is either for engineers or for hype. There isn't much for someone who's curious and starting from zero. I'm close to zero too. I'm a few steps ahead at most, and happy to say so.
I'll be skeptical of demos. A great video isn't a shipped product. I'll keep asking whether it was sped up, remote-controlled, or the one good take out of fifty.
That's it. Below is the guide I wish I'd had on day one.
Part 2: The beginner's guide
The one idea that connects them all
Every robot does three things over and over:
- Sense: figure out what's going on (where am I, what's around me, how are my joints bent).
- Decide: pick what to do next.
- Act: spin a motor, move a leg, close a hand.
A remote-controlled machine hands the "decide" step to a person. Modern robots try to make that decision themselves, hundreds of times a second.
The easiest way I've found to keep the parts straight is to compare them to your own body:
| On you | On a robot | What it's doing |
|---|---|---|
| Muscles | Motors and joint actuators | Turning energy into motion |
| Eyes | Cameras, depth sensors, lidar | Seeing the world |
| Inner ear | IMU (an accelerometer and gyroscope) | Knowing which way is up |
| Reflexes | Control software | Turning goals into motor commands |
| Hands | Hands, grippers, tools | Doing the actual job |
| Food | Batteries | Still the thing everyone wishes were better |
The kinds of robot you'll keep seeing
Robot arms
The simplest place to see the loop is a robot arm bolted to a table. It doesn't have to balance or walk. It only has to see an object, decide how to reach it, and move its joints to get there. That makes arms a common starting point for AI robotics, and the cheapest way to try it yourself (more on that below).
Quadrupeds: robot dogs
Four legs are easier than two. A dog-shaped robot can stand still without falling over, which buys a lot of forgiveness.
Boston Dynamics' Spot is the famous one. It sees with five stereo depth-camera pairs: two in front, one on each side, and one in back (Boston Dynamics). At the cheap end, Unitree's online store lists the Go2 Air from $1,600 before shipping, as of October 31, 2026 (Unitree store).
MIT's Mini Cheetah, the backflipper, is the one to study. It weighs about 20 lb (9 kg) (MIT; ICRA 2019 paper). Each of its 12 motors is "about the size of a Mason jar lid." Those motors come up again below.
Humanoids: robots shaped like us
Humanoids are the hardest and the most hyped. Why build a robot shaped like a person? Because the world is already built for people: stairs, door handles, shelves, tools. In theory, a humanoid can work in a warehouse without the warehouse changing.
A few you'll see in the news:
- Atlas (Boston Dynamics). The old hydraulic Atlas, the one that did parkour, retired in April 2024. An all-electric version replaced it, aimed at real work starting with Hyundai's factories (Boston Dynamics).
- Digit (Agility Robotics). One of the few humanoids doing paid commercial work, including with logistics company GXO. Agility expects to start trading on Nasdaq on November 5, 2026 (Agility; our story).
- Unitree G1. A small humanoid you can actually buy. Unitree's store lists it at $13,500 before shipping, as of October 31, 2026 (Unitree store).
- Meta's "Max" (reported). The Information reports that Meta built a humanoid prototype with Chinese robot maker Fourier. Meta says it's internal research and not for sale (our story).
Real or hype? Three questions for any robot video
| Ask | Why it matters | Where to look |
|---|---|---|
| Is it sped up? | 2x footage makes slow robots look capable | Caption, on-screen "2x" or "4x" label |
| Is someone teleoperating it? | A person off camera may be making the decisions | Caption, fine print, the company's blog post |
| Is it doing the task once or all day? | One clean take can hide dozens of failures | Whether the company reports hours, shifts, or paying customers |
How they actually work
Actuators: the muscles
On a toy car, a motor spins a wheel, and that's mostly it. On a walking robot, every joint needs a motor that can push and be pushed back.
Here's how the Mini Cheetah's motors work. Each one is a brushless motor with a small 6:1 gearbox (MIT). That's a deliberately low gear ratio. A big gear reduction gives you lots of torque but makes the joint stiff, like trying to turn a car's wheels with the engine off. A low ratio keeps the joint "backdrivable." When the foot slams into the ground, the motor gives a little instead of breaking, and the robot can feel the impact through the motor current. MIT's researchers called this proprioceptive actuation (MIT Cheetah actuator paper). It's often called quasi-direct drive.
That answered my backflip question. Part of how the robot knows where the ground is: it feels it through its own motors.
Older robots like the original Atlas used hydraulics, which are very strong but messy and complex. The industry has moved toward electric motors, and Atlas itself went electric in 2024.
Sensors: the senses
Robots need two kinds of sense:
- Looking outward: cameras, depth cameras, and lidar (a laser rangefinder) to map the room and spot obstacles.
- Looking inward: encoders that measure each joint's angle, and an IMU that measures tilt and acceleration. Your inner ear does the IMU's job for you.
The inward senses are underrated. A walking robot depends on knowing exactly how its own body is moving, often more than on what its cameras see.
Control: the reflexes
Control is the software that turns "walk forward" into thousands of tiny motor commands. The classic approach is to model the physics and solve for the best move many times a second. Using a method called model-predictive control, the Mini Cheetah ran at up to 2.45 m/s (ICRA 2019 paper).
The newer approach is to let the robot learn. In 2021, researchers at ETH Zurich simulated thousands of robot dogs in parallel on a single graphics card. Their quadruped learned to walk on flat ground in under four minutes of training and on rough terrain in about twenty, and the learned controller was then moved onto the real robot (Rudin et al., CoRL 2021). That's called sim-to-real, and it's a big reason legged robots improved so quickly.
AI: the part that's changing fastest
Walking is a reflex. Knowing what to do, like "put the apple in the bowl," is a different problem.
In 2023, Google DeepMind showed RT-2, a model that takes camera images and a plain-language instruction and outputs robot actions directly. It does this by treating motor commands like words in a sentence (Google DeepMind). Models like this are called vision-language-action models, and many humanoid companies are betting on some version of the idea.
A rough way to think about it: classic control keeps the robot upright, learned policies give it athletic skills, and big AI models are starting to give it common sense. That last layer is the least finished.
Where to start
You don't need a lab. Pick the path that matches how you like to learn.
If you learn by building, build an arm.
- Hugging Face's open-source LeRobot project works with the SO-101, a 3D-printable arm. At launch, TechCrunch reported it started at about $100 (TechCrunch). The design files are on GitHub. You can teach it a task by demonstration and train a small policy, the same basic loop the big labs use.
If you want to watch and read first:
- Boston Dynamics on YouTube to see legged robots straight from the source. Keep the three questions above handy.
- Russ Tedrake's free MIT course notes, Underactuated Robotics, when you're ready for the math of walking.
FAQ
What's the difference between a robot and a remote-controlled machine? Both sense and act. The difference is who does the deciding. A remote-controlled machine leaves it to a person. A robot makes at least some of those calls itself.
Why build robots shaped like people? Because homes, warehouses and factories are already built for human bodies: stairs, door handles, shelves, tools. A humanoid could, in theory, work there without anything being rebuilt.
What's the cheapest robot dog or humanoid you can buy? As of October 31, 2026, Unitree's store lists the Go2 Air robot dog from $1,600 and the G1 humanoid at $13,500, both before shipping.
How do robots learn to walk? Many now train in simulation first, with thousands of virtual copies practicing in parallel, and then transfer what they learned to the real robot. This is called sim-to-real.
What I'm still figuring out
I don't know yet whether humanoids will be in homes in five years or twenty, and I haven't found anyone who can answer that honestly. What I do know is that the distance from watching a robot do a backflip to understanding how it keeps its balance is shorter than it looks.
If you're starting from the same place, I'd love to learn alongside you. And if I got something wrong here, tell me.
Last updated October 31, 2026. Prices are list prices on the date shown and change often.
— Nate