A beginner's guide to robotics
A beginner's introduction to what a robot is, its types, what robots can and cannot do today, the opportunities, and how to get your first one running.
A robot is a machine that senses the physical world and decides what to do about what it sensed and then acts back on that world without a human steering every move. Everything else in robotics exists to make that loop run fast and reliably.
Somewhere in San Francisco right now a car with an empty driver's seat is pulling over to pick up a passenger, and the company behind it does that half a million times a week across all the cities it runs in, with a million as the target before 2026 ends. That is the shift in robotics everyone can see. The quieter one is for engineers, who can now buy a robot and write software for it instead of waiting on a lab to let them near one.
In this article I want to map out that shift, starting with what a robot is and ending with one running on your laptop.
A robot vs. a machine
The definition in the box above sounds almost too simple until you hold it against the machines around you. A dishwasher acts on the physical world but never senses anything, a security camera senses constantly but never acts on what it sees.
A robot does both, in one loop: sense, plan, act. The robot reads its sensors -> plans a response -> executes it -> reads the sensors again to find out what actually changed. In a robot this loop runs over the physical world, over positions, obstacles, and contact, and all of those change every time the body moves.
InteractiveDrag the grey box and see what happens.
The types of robots
INDUSTRIAL ARM The oldest and most numerous type. A bolted down machine that has been welding and painting and placing parts since the 1960s, with more than four and a half million operating in factories worldwide and China alone running past two million.
WAREHOUSE ROBOT The same sense plan act loop on wheels instead of a bolted base. Mobile robots carry shelves and totes across fulfillment centers by the hundreds of thousands, navigating around each other and around people all shift long.
QUADRUPED Legs came after wheels because legs handle stairs and rubble and curbs that wheels cannot. Robot dogs like the Unitree Go2 walk patrols on construction sites and substations, and the base configuration sells for 1600 dollars.
HUMANOID The newest arrival and the most ambitious one, since the entire built environment assumes a human shape and every door handle and shelf height and staircase was designed for a body like ours. Machines like the Unitree G1 now put that form factor on the market at 13500 dollars.
DRONE Swap legs for propellers and you get the drone, which counts as a robot by the definition above whenever it flies a survey or an inspection without a pilot micromanaging every second of the flight.
The subsystems that make a robot work
Every robot regardless of shape is built from the same handful of subsystems, and once you can name them you can watch any robotics demo and understand what is actually happening on screen. Pin them to the loop to keep them straight. Perception and localization do the sensing; navigation, manipulation, and reasoning do the planning; and control does the acting.
InteractiveDrag the cups, pose the arm, or .
Every part of the figure above is one step in sense, plan, act. PERCEPTION is sense: the cameras turn what is in front of the robot into something it can use, a cup on the floor, a person blocking the path. In practice that is a detection model like YOLO running about thirty times a second. That is what the dog is doing when it looks at the cups before it moves.
LOCALIZATION is still sense, but about the robot itself. There is no GPS indoors, so the dog has to know where it is on the floor while it is walking. It builds a map of the room and keeps updating its place on that map as it goes. Roboticists call this SLAM, building the map and locating yourself inside it at the same time.
Once it knows what is around it and where it is, NAVIGATION plans the walk. It picks a path to the cup and keeps redrawing that path when something moves, which is why it can walk around a cup you dragged into the way.
MANIPULATION is the arm. Reaching for a cup means figuring out how every joint should bend so the gripper gets there without knocking anything over. Harder than walking, because now the robot is reasoning about contact, not just empty space. On a humanoid like the Unitree G1, one approach is to sample around two million arm poses per arm so the robot knows what it can reach before trying a motion.
Under all of that runs CONTROL, the fastest loop in the stack. Navigation decides where to go, control decides what every motor has to do to make that happen. It turns a step into torque commands for each of the twelve leg joints, hundreds of times a second, and corrects for every slip and wobble in between, a push, a slick patch of floor, the weight of the arm shifting as it reaches. It knows nothing about cups or tables and only keeps the hardware upright and the motion smooth while the slower loops above it think.
On top sits REASONING, the newest piece. Instead of programming each step yourself, you give the robot a goal in plain language and it breaks that into the skills above: find the cup, walk over, pick it up, put it on the table.
What robots can actually do today
The honest 2026 picture is robots are superhuman inside structured environments and clumsy outside. A robot is as good as or even better than humans in a factory or a warehouse aisle, the world stays predictable enough and the robot rarely meets anything unexpected. A factory arm welds the same seam millions of times, warehouse robots slide shelves past each other all shift without colliding, and the robotaxis in San Francisco run paid rides all day on streets they have mapped down to the lane marking.
Step outside structure and everything changes fast, most of robotics in 2026 still lives in that gap, between a demo that works and an ordinary home. Having said that, the gap will not be bridged by a single breakthrough but by developers taking robots working in demos and making them survive one real messy environment at a time, and the models improving with every attempt since each run produces the interaction data they train on. You do not need a lab to start poking at it.
Get your first robot running
Getting started in robotics is cheaper than it used to be. MuJoCo is a free physics engine for testing a robot on your laptop before you put it on hardware. A robot arm is about a hundred dollars in parts. So, let's get your first robot running right now in MuJoCo. Later on, pick a robot and a small task, get it working in simulation, and write up what broke or didn't work along the way.
Set up Python and install MuJoCo:
uv venv --python "3.12" source .venv/bin/activate uv pip install mujoco
Then open the interactive simulator. No robot needed. Drag any robot model from MuJoCo Menagerie, like the Unitree Go2, into the window:
python -m mujoco.viewer
What robots will be able to do next
The bet is the same one that made language models work: train on enough data and the system starts handling situations you never programmed by hand. Vision-language-action models like π0 and GR00T already run one network across different robots and different tasks, and every major lab is racing to collect the physical interaction data those models need. My bet is that within a decade you will describe a task to a robot the way you describe one to a coding agent today, and the robot will handle the sensing and planning and acting on its own.
The opportunity if you are coming from software
If that bet pays off, someone still has to connect the model to the motors. The hardware is cheap enough to buy, the models are getting cheaper to run. What every team rebuilds from scratch is the software in between, and that is the part most software engineers already know how to write.
A robot stack today looks like any other backend: Python, message streams, APIs, tests, deployment. Your first program might subscribe to a camera feed and publish a velocity command in a dozen lines of Python. The jobs opening up are in perception, simulation, fleet tooling, and the data pipelines feeding the model labs.
Getting hands-on is also cheaper than it has ever been, as mentioned earlier, since simulators like MuJoCo are free and an open source arm from the LeRobot project costs around a hundred dollars in parts. The way into robotics looks like the way people have always gotten into software: picking one small project and building it in the open, running a robot in simulation until it does one real task end to end and then writing about everything that broke along the way.