Look! A robot! π€
Robots are machines that can move. Some robots roll. Some robots walk.
A robot can help you. It can pick things up!
Robots do not feel. But they are fun to watch. β¨
What Is a Robot?
A robot is a machine that can move and do jobs. It follows steps that a person gives it. Those steps are called a program.
How Does a Robot Know Things?
Some robots have cameras for eyes. Here is a new word: sensor. A sensor is a part that feels or sees things. Your skin is like a sensor. It feels hugs!
Can Robots Think?
Robots do not think the way you do. They follow their program step by step. The program tells the robot what to do first, what to do next, and what to do last. That is how a robot cleans a floor or builds a car part.
What Makes a Robot a Robot?
Almost every robot has three parts that work as a team. First, sensors gather information about the world, like cameras, touch sensors, and microphones. Second, a computer brain reads that information and decides what to do. Third, actuators are the moving parts, like motors and wheels, that carry out the decision.
The three parts run in a loop: sense, think, act. A vacuum robot senses a wall, decides to turn, and its wheels act. Then it starts sensing again. This loop can repeat hundreds of times every second.
Robots You Might Already Know
Robot arms weld and paint cars in factories. They never get tired, and they can lift parts that weigh more than a person. Warehouse robots roll along the floor carrying shelves to the workers who pack boxes.
At home, robot vacuums map your rooms and clean while you are at school. Lawn robots mow grass. And on Mars, NASA rovers like Curiosity and Perseverance drive around collecting rocks. No human has ever set foot on Mars, but robots have been working there since 1997.
Do Robots Feel Anything?
A robot can act surprised, but it is not surprised. Programmers give robots faces and voices because people find it easier to work with something friendly. Underneath, the robot is still running its sense-think-act loop. It follows the program, even when it smiles at you.
The Sense-Think-Act Loop
Every robot, from a factory arm to a Mars rover, runs the same loop. Sensors report the state of the world. A controller compares that state to a goal. Actuators push the world closer to the goal. Then the loop starts over, often hundreds of times per second. Fast loops make robots responsive; slow loops make them clumsy.
The controller in the middle is where the interesting problems live. A motor command is never perfect. Gears have friction, batteries sag, floors are uneven, and a gust of wind shoves a drone sideways. Control theory is the mathematics of commanding imperfect machines anyway, and it is the reason robots can walk, fly, and balance at all.
Degrees of Freedom
Engineers describe a robot's mobility with a single number: its degrees of freedom, which count the independent ways it can move. A point on a screen has 2 (left-right and up-down). Your arm has 7 (three at the shoulder, one at the elbow, three at the wrist), which is why you can hold your hand still while moving your elbow, a trick called redundancy.
More degrees of freedom mean more flexibility but much harder math. A 6-joint robot arm needs to solve for 6 angles every time it reaches for something, and there can be several valid answers. Picking the wrong one makes the arm swing through a table. Motion-planning algorithms search for a safe, smooth path before the robot moves.
Sensors and Sensor Fusion
No single sensor tells the truth. Cameras fail in the dark, GPS fails indoors, and wheel odometry drifts as tires slip. Good robots combine several imperfect sensors, a technique called sensor fusion. A delivery robot might use GPS outdoors, cameras for lane lines, and lidar (laser distance sensing) to dodge obstacles. When the sensors disagree, the math weighs each by its reliability.
The Kalman filter, published by Rudolf Kalman in 1960, is the classic fusion algorithm. It keeps a running estimate of the robot's state and updates it every time a new measurement arrives, trusting each measurement exactly as much as it deserves. Kalman filters guided Apollo to the Moon and still sit inside most drones and self-driving systems today.
Why Walking Is Hard
Here is a strange fact about robotics: the things that are easy for humans are hard for robots, and vice versa. A robot can weld the same car door ten thousand times without a mistake, but teaching it to walk across a cluttered room took decades. This mismatch even has a name: Moravec's paradox, after AI researcher Hans Moravec, who observed that high-level reasoning needs little computation while low-level sensorimotor skills need enormous amounts of it.
The reason is evolutionary time. Walking, balancing, and recognizing objects are skills your brain refined over hundreds of millions of years, so they feel effortless. Multiplying numbers is new, so it feels hard. Robots inherited the opposite difficulty curve: arithmetic is trivial for a processor, but staying upright on two legs requires solving physics problems in real time.
A Short History of Machines That Act
The word robot comes from the Czech word robota, meaning forced labor, coined by Karel Capek in his 1920 play R.U.R. The machines, however, arrived later and far more quietly. In 1961, George Devol's patent (US2988237, "Programmed Article Transfer") became the first industrial robot, the Unimate, installed at a General Motors plant in New Jersey to handle hot die-cast parts. It had no sensors at all. It simply repeated a recorded sequence of motions, open-loop, with hydraulic brute force.
Real intelligence arrived in layers. From 1966 to 1972, SRI International built Shakey, the first mobile robot that could perceive its environment and plan. Shakey combined a TV camera, range finders, and bump sensors with the STRIPS planner, which broke goals into subgoals and reasoned about the effects of actions before taking them. Shakey was slow and clumsy by modern standards, but it proved the sense-plan-act architecture that organized robotics for decades.
Control Theory: Commanding Imperfect Machines
The workhorse of robot control is the PID controller, which computes a correction from three terms: how far off you are now (proportional), how long you have been off (integral), and how fast the error is changing (derivative). A quadcopter holding altitude runs this loop thousands of times per second.
Each term has a personality. The proportional term reacts to the present error but cannot quite finish the job. The integral term remembers past errors and eliminates the leftover steady-state offset, though it can overshoot. The derivative term anticipates the future by watching the error's trend, damping oscillations. Tuning the three gains, Kp, Ki, and Kd, is half science and half craft, and mis-tuned gains are why cheap drones wobble.
For arms, the physics gets heavier. Each joint must produce torque against gravity, inertia, and the motion of every other joint. The required torque follows from rigid-body dynamics: torque equals the arm's inertia times its angular acceleration, plus the torques needed to fight gravity and to counteract the other moving links. In the simplest form, the relationship between force and rotation is:
where Ο is torque, r is the distance from the joint, and F is the applied force. This is why the motor at a robot's shoulder must be far stronger than the motor at its wrist: the shoulder carries the leverage of the entire arm.
Learning Instead of Programming
Classical robotics required engineers to model everything: the arm's mass, the friction, the camera geometry. Modern robots increasingly learn from data instead. Reinforcement learning lets a simulated robot try millions of grasps in hours, then transfer the skill to the real machine, a pipeline called sim-to-real. The gap between simulation and reality remains the central headache: policies that work perfectly in a physics engine can fail the moment a real cable snags.
Large language models added a new layer on top. A household robot given the command "clean up the spilled cereal" needs to decompose it into find the box, pick it up, get a broom, sweep, dispose, a plan that requires common sense no control loop contains. Researchers now pair LLMs as high-level planners with classical controllers as low-level executors, letting the language model decide what to do while the PID loops handle how. It works surprisingly well and fails surprisingly often, usually at the boundary where a reasonable-sounding plan meets a physics the model never learned.
Safety and the Three Laws
Isaac Asimov's Three Laws of Robotics (1942) are fiction, but they shaped real thinking about robot safety. The actual standards are less elegant and more useful: ISO 10218 governs industrial robot safety, and ISO/TS 15066 sets force and pressure limits for collaborative robots that work beside humans, down to how hard a robot may press against different body parts before it must stop.
The real safety questions are messier than the laws. Who is responsible when an autonomous system makes a harmful decision: the programmer, the operator, or the owner? How do you test a system that learns after deployment? And what happens to the labor market when the sense-think-act loop gets cheap enough to put in every warehouse? These are the questions the adult level takes up next.
From Die-Casting to Dexterity: What the History Actually Shows
The standard story of robotics is a march from dumb machines to smart ones. The history is more instructive than that. George Devol filed his "Programmed Article Transfer" patent in 1954 and received US2988237 in 1961; the first Unimate went to work at General Motors' Inland Fisher Guide Plant in Ewing Township, New Jersey, handling hot die-cast parts. It was not intelligent, but it was repeatable, and repeatability turned out to be the commercially decisive property. For two decades, that was the entire industry: open-loop arms doing dangerous, dirty, dull work in cages, separated from humans by fences and interlocks.
The intelligence arrived in research labs first. Shakey at SRI (1966-1972) demonstrated that a mobile robot could perceive, plan with STRIPS, and act, though at speeds that made it a curiosity. The Stanford Arm (1969, Victor Scheinman) gave researchers an electrically driven, computer-controlled manipulator to experiment with. Scheinman later designed the PUMA arm, which became the standard industrial research platform. Honda's humanoid program, beginning with the E0 in 1986 and culminating in ASIMO's 2000 debut, attacked the hardest problem, bipedal walking, and proved it solvable with enough engineering. Boston Dynamics, founded by Marc Raibert in 1992 after his Leg Lab work at CMU and MIT, turned dynamic legged locomotion from a research demo into products: BigDog (2005), then Spot (commercially available 2020) and the Atlas humanoid.
Notice what actually crossed into the mass market. Not the humanoids: the Roomba (iRobot, 2002) succeeded because it solved a narrow problem with cheap sensors and a behavior-based architecture descended directly from Brooks's subsumption ideas. Industrial arms succeeded because cages made safety a non-problem. The pattern: robots commercialize where the environment can be constrained, not where the robot is most impressive.
Moravec's Paradox and Where the Difficulty Really Lies
Hans Moravec's observation, articulated in Mind Children (1988), remains the best single lens on the field: tasks that feel hard to humans (chess, calculus, logic) require little computation, while tasks that feel effortless (seeing, walking, grasping) demand enormous computation. The reason is that a billion years of evolution optimized sensorimotor competence into neural hardware we mistake for simplicity.
This has a direct consequence for anyone forecasting automation. The jobs that look cognitively demanding, document review, scheduling, basic coding, are the ones current systems handle best. The jobs that look simple, plumbing, elder care, warehouse picking in cluttered bins, remain stubbornly hard. The paradox predicts exactly this inversion, and the labor data has largely cooperated: automation has bitten hardest at routine cognitive work while physical dexterity in unstructured environments remains the bottleneck.
That bottleneck is narrowing, though not evenly. The dexterity gap is closing fastest where the environment is structured (factories, warehouses with standardized bins) and slowest where it is not (homes, construction sites, hospitals). Anyone evaluating a robotics claim should ask first: how structured is the environment it was demonstrated in?
Scale, Economics, and the Honest Numbers
The International Federation of Robotics' World Robotics 2024 report counted roughly 4.28 million industrial robots in operation worldwide, with about 541,000 new installations in 2023. South Korea leads in robot density at about 1,012 robots per 10,000 manufacturing employees. These are large numbers that describe a mature industry, and they are almost entirely about fixed arms in structured environments.
The honest limitations deserve equal weight. Dexterous manipulation in unstructured environments is still a research problem; the celebrated demos (a robot folding laundry, cooking a meal) typically run under controlled conditions with significant human preparation. Battery energy density constrains mobile robots far more than computation does: a humanoid that could work a full shift needs an energy system we do not have at acceptable weight. And the unit economics of humanoid robots remain unproven at scale; the announced prices from Tesla (Optimus), Figure, and others are targets, not demonstrated costs.
"It is comparatively easy to make computers exhibit adult-level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility."
Hans Moravec, Mind Children (1988)
Safety, Responsibility, and the Questions Without Clean Answers
Asimov's Three Laws were a literary device, and treating them as an engineering specification would be a category error. The real safety regime is standards-based and prosaic: ISO 10218 for industrial systems, ISO/TS 15066 for collaborative robots, with force and pressure limits calibrated to specific body regions. These standards assume the robot's behavior is specified in advance.
Learning systems break that assumption. A policy trained by reinforcement learning is a statistical artifact, not a specified behavior, and verifying it the way we verify a PID loop is an unsolved problem. The industry's working answer is layered: keep the learning at the planning level, keep classical controllers with provable properties at the execution level, and wrap the whole thing in monitored cages or speed-and-separation monitoring. It is defense in depth, and like all defense in depth, it manages risk rather than eliminating it.
The labor questions are the ones parents ask about, and they deserve candor rather than comfort. Industrial robots have displaced specific categories of manufacturing labor while creating others (programming, maintenance, integration). The net employment effect in manufacturing-heavy regions has been negative in several careful studies and roughly neutral in others; the honest summary is that the effect is real, uneven, and mediated by policy. What is new is not displacement itself but its direction: the current wave targets routine cognitive work first, which reverses the historical pattern where machines took the physical jobs. That reversal is the thing to watch, because our institutions were built for the old pattern.
Sources
- Devol, George C. "Programmed Article Transfer." U.S. Patent 2,988,237, filed 1954, issued 1961.
- Nilsson, Nils J. "Shakey the Robot." SRI International Artificial Intelligence Center, Technical Note 323, 1984.
- Brooks, Rodney A. "A Robust Layered Control System for a Mobile Robot." IEEE Journal on Robotics and Automation, 1986.
- Moravec, Hans. Mind Children: The Future of Robot and Human Intelligence. Harvard University Press, 1988.
- International Federation of Robotics. World Robotics 2024: Industrial Robots. Frankfurt, 2024.
- ISO 10218-1:2011 and ISO 10218-2:2011, Safety requirements for industrial robots; ISO/TS 15066:2016, Collaborative robots.
- Kalman, R.E. "A New Approach to Linear Filtering and Prediction Problems." Journal of Basic Engineering, 1960.