Robotics

Robotics is the engineering discipline concerned with machines that sense their surroundings and act on them physically. The field draws on mechanical design, control theory, electronics, and computer science, and its products range from a welding arm bolted to a factory floor to a surgical instrument, a warehouse robot, and a legged machine that climbs stairs 1.

The word robot comes from Karel Čapek's 1920 play R.U.R., and the first industrial robot, Unimate, began work at General Motors in 1961. Robots are distinguished from other machines less by appearance than by programmability and feedback: they measure what they are doing and adjust. That capacity is what allows a robot to work on a moving assembly line, or to pick an object it has not seen before 2.

This entry covers the history of the field, the anatomy of a robot, sensing and control, the main categories, the economics of adoption, and the questions of labour, safety, and accountability that accompany their spread.

The word robot entered English through Čapek's play, in which artificial workers revolt against their makers. The engineering reality arrived later and more prosaically. George Devol patented a programmable manipulator in the 1950s, and Joseph Engelberger's company, Unimation, installed the first Unimate at a General Motors plant in Ewing Township, New Jersey, in 1961, where it moved hot die-cast parts 2.

Industrial robotics grew through the 1970s and 1980s as carmakers automated welding, painting, and assembly. The articulated six-axis arm became the standard form, supported by hydraulic or electric actuators and increasingly by computer control 1.

The 1990s and 2000s pushed robots out of the factory. Surgical systems, vacuum cleaners, warehouse carriers, and autonomous vehicles appeared, along with humanoid research platforms such as Honda's ASIMO. Each required better perception and safer interaction, and each exposed the gap between a structured factory and an unstructured world. The 2010s added machine learning: policies trained in simulation began to transfer to hardware, and legged robots learned to run and recover from falls 3.

A robot consists of a mechanical structure, actuators, sensors, a controller, and a power source. In an industrial arm, the structure is a chain of links joined by rotary joints; each joint adds a degree of freedom, and six are enough to place and orient an object in three-dimensional space 4.

Actuators convert energy into motion. Electric servomotors dominate because they are precise and easy to control; hydraulic actuation offers greater force density and was long used in legged machines; pneumatic actuators are cheap and fast but hard to position accurately. The choice determines not only the robot's strength but its stiffness, its efficiency, and how safely it can behave near people.

The end effector is the tool at the end of the chain: a gripper, a welding torch, a camera, or a scalpel. Much of a robot's usefulness lies in this attachment rather than in the arm itself. The controller coordinates the joints, reads the sensors, and enforces the limits that keep the machine and the people around it safe 4.

Safety is a design property, not an add-on. Industrial standards require risk assessment, guarding or protective stops, and procedures for maintenance and teaching, when a person may be inside the robot's workspace. The 2025 revision of ISO 10218 updates requirements for functional safety and collaborative operation 4.

Proprioceptive sensors report the robot's own state: encoders measure joint angles, torque sensors measure force, inertial units measure acceleration and orientation. Exteroceptive sensors report the world: cameras, lidar, tactile skins, and microphones. A mobile robot combines the two to estimate where it is while it moves, a problem known as simultaneous localisation and mapping 5.

Control closes the loop between them. A controller compares the measured state with the desired one and computes a correction, using a proportional-integral-derivative law for simple joints and model-based methods for whole-body motion. Planning operates above control, converting a task into a trajectory that respects the machine's kinematics and its environment.

Model-based control works well when the robot's dynamics are known and the environment is predictable. It struggles with contact, deformation, and terrain that cannot be modelled precisely. Learning-based control offers an alternative: a neural network is trained in simulation to map sensor readings to actions, and the resulting policy is transferred to hardware. The ANYmal quadruped learned to follow velocity commands and recover from falls this way, though the transfer depends on accurate simulation and careful calibration 3.

The two approaches are converging. Modern humanoids use model-based controllers for balance and learned policies for locomotion and manipulation, with a supervisory layer that decides which to trust 6.

Industrial robotics is the oldest and largest market. The Unimate performed die-casting work that was hot, heavy, and dangerous, and welding, painting, palletising, and assembly followed. The articulated six-axis arm became the standard form, with specialised geometries where speed or workspace mattered: SCARA arms for flat assembly, delta robots for high-speed picking 2.

The virtue of an industrial robot is repeatability, measured in fractions of a millimetre over millions of cycles, rather than adaptability. A welding robot repeats a programmed path with a consistency no human can match, and it does so without fatigue. The cost is setup: fixtures, programming, and safety fencing are often more expensive than the arm itself, which is why automation favours high-volume production 4.

The market is concentrated. In 2024, 542,000 industrial robots were installed worldwide, the fourth consecutive year above 500,000, and the operational stock reached 4,664,000 units. Asia accounted for 74 per cent of new deployments, Europe 16 per cent, and the Americas 9 per cent; China alone took 54 per cent, with 295,000 installations and a stock above two million 1.

Collaborative robots changed the safety model. Instead of fencing a machine off, a cobot is designed to limit force and speed so that it can work alongside people. That makes it easier to deploy in small cells, but it also reduces payload and speed, and the safety case depends on the application rather than the robot alone 4.

Service robots work outside the factory: vacuuming floors, delivering meals, moving goods in warehouses, inspecting pipelines, and assisting surgery. Field robots work outdoors, in agriculture, mining, construction, and defence, where terrain and weather make perception difficult 1.

Some of these are mature businesses. Surgical systems that translate a surgeon's hand movements have become standard for certain procedures; warehouse robots that carry shelves to pickers reshaped logistics; and autonomous trains and mine haulers operate in controlled environments. Others remain demonstrations, particularly those that require manipulation in cluttered spaces or interaction with the public.

The pattern is that success depends less on the robot than on the environment. A structured setting with reliable power, maintenance, and clear rules is far easier than a public street or a hospital corridor. The robot that works well in a warehouse often fails in a home, not because its hardware is worse but because the home was never designed for it 1.

Field robots add regulatory complexity. A drone, an autonomous tractor, and a surgical device are governed by different rules, and the safety case must account for weather, bystanders, and failure modes that a factory never sees 7.

Humanoid robots are the field's most visible and least settled branch. Honda's ASIMO, unveiled in 2000, walked and climbed stairs but was slow and expensive; Boston Dynamics' Atlas, first shown in 2013, demonstrated dynamic balance and acrobatics that changed expectations about what legged machines could do. In 2024 Boston Dynamics replaced the hydraulic Atlas with an all-electric design intended for commercial work 6.

The modern approach combines classical control with learned policies. Simulated environments allow millions of trial runs, and policies trained in simulation can transfer to hardware with careful modelling. Whole-body controllers now handle uneven ground and unexpected pushes, and reinforcement learning has produced running, jumping, and recovery behaviours that would be difficult to hand-code 3.

The humanoid form has a practical argument: factories, warehouses, and homes are built for human bodies, so a machine that fits those spaces can use existing tools and infrastructure. It also has a cost: legs are less stable and less efficient than wheels, hands with human-level dexterity remain difficult to build, and battery life is measured in hours. Whether general-purpose humanoids make commercial sense outside a few niches is unresolved 6.

The commercial interest is real. Several companies have announced humanoid pilots in manufacturing and logistics, often with limited disclosure of hours worked, failure rates, or cost per task. Those claims should be treated as claims until independent data exists, because the gap between a demonstration and a productive shift is where most robotics programmes fail 6.

The economics of a robot are the economics of replacing a task. A robot costs capital, installation, programming, maintenance, and energy; it saves labour, improves consistency, and can run shifts that people cannot. The payback period depends on utilisation: a robot that runs for one shift may never repay its cost, while the same robot running around the clock can repay it in a year or two 1.

Adoption is uneven because the fixed costs fall hardest on small firms. A large carmaker can justify bespoke tooling and a team of integrators; a small machine shop cannot. That is one reason robot density varies enormously between countries and industries, and why leasing, robots-as-a-service, and simpler programming interfaces have become important. The direction of travel is towards lower integration cost, not necessarily lower hardware cost.

The market data reflect these incentives. Industrial installations have held above half a million units a year, but growth has slowed, and the stock is concentrated in a few manufacturing economies. China's dominance reflects both its manufacturing base and deliberate industrial policy; Europe's installations fell 8 per cent in 2024, while Asia grew. Service robotics is smaller and harder to measure, because the category includes inexpensive consumer devices and expensive medical systems 1.

Robotics also interacts with energy and supply chains. Robots can reduce waste and energy use in production and can support reshoring. Those benefits are difficult to attribute, and they are usually claimed by vendors rather than measured independently 7.

Automation has always displaced particular tasks, and the evidence on net employment is mixed. Acemoglu and Restrepo estimated that one more industrial robot per thousand workers reduced the US employment-to-population ratio by about 0.2 percentage points and wages by about 0.37 per cent, with effects concentrated in routine manual occupations. That is a measurable local effect, not a verdict on the total number of jobs 8.

Robots also create demand for maintenance, programming, integration, and supervision, and they can raise productivity enough to lower prices and increase output. The distributional effects are uneven: workers whose skills are specific to the automated task bear the cost, while consumers and firms capture much of the benefit 8.

Safety is regulated but imperfect. Industrial standards require risk assessment, guarding, and safe maintenance procedures, and OSHA documents the hazards of robotic systems, including unexpected motion during teaching and maintenance. The consequences of a mistake are severe because robots are strong and fast 5.

Accountability for accidents is legally unsettled. When a robot injures someone, liability may fall on the manufacturer, the integrator, the operator, or the software supplier, and existing product-liability rules map imperfectly onto systems that learn and are updated remotely. The European Union's machinery regulation brings software and AI-enabled safety functions into scope, but it does not resolve the underlying question of who is responsible when behaviour emerges from training rather than design 7.

The labour debate is the oldest disagreement. Industry bodies emphasise productivity, quality, and the jobs created in robot manufacturing and maintenance; labour economists emphasise displacement in routine occupations and the uneven distribution of gains. Local displacement can coexist with aggregate employment growth 8.

Humanoids attract the widest gap between expectation and evidence. Vendors present general-purpose humanoids as a platform that will eventually do many tasks; sceptics note that purpose-built machines are cheaper, faster, and safer for almost every specific job. The resolution depends on whether manipulation and mobility improve enough to justify a general-purpose machine, which is an empirical question that current deployments do not answer 6.

Safety has its own divide. Collaborative robots are marketed as inherently safe, but their safety depends on force limits, speed monitoring, and the application; a cobot wielding a knife is not safe because it is a cobot. Standards are written for machines whose behaviour is defined by design, and they struggle with learned policies whose failure modes are not enumerable in advance 4.

Finally, there is disagreement about what counts as a robot. A software agent that books a flight is not a robot in the industrial sense, but the boundary is blurring as machines use language models for instruction and planning. The term is likely to remain a family resemblance rather than a strict category, and arguments about whether a given system is truly a robot are often arguments about marketing rather than engineering 1.