Robotics
in a Hurry

The Whole Thing in One Page

The public picture of a robot is a metal person. The machines doing most robotic work are less theatrical: arms welding car bodies behind fences, mobile platforms carrying shelves, grippers sorting parcels, systems steering surgical instruments, crawlers inspecting pipes, milking machines working around cattle and rovers choosing paths across Mars. A robot is better understood as a bargain between a task, a body and a world.

Start with the task. Engineers must decide what counts as success, what variation the machine must tolerate, what can be fixed in advance and what failure would cost. Much of robotics consists of placing intelligence outside the robot. A fixture aligns the part. Floor markings simplify navigation. A tray presents every component in the same pose. A human handles exceptions. The machine succeeds because the whole environment has been designed as carefully as the machine.

Then the body. Wheels are efficient on smooth floors. Legs trade efficiency for access. A fixed arm gains stiffness by giving up mobility. A soft gripper accepts less precision in exchange for safer contact. Motors, gears, hydraulics and pneumatics turn stored energy into force, but each brings heat, friction, backlash, weight and limits. The form of a robot is a decision about which physics to accept.

Sensors do not hand over reality. They produce measurements: pixels, ranges, rotations, accelerations, currents and forces. Those measurements arrive noisy, delayed and from different coordinate frames. The robot must estimate what is happening, including its own pose. Calibration, sensor fusion, localisation and mapping are therefore central, not supporting details.

Geometry translates an intended movement into joint motion. Kinematics says where a mechanism can go. Dynamics says what force and torque are needed to get there. Planning searches for a feasible path. Control makes the physical machine follow it, measures the error and corrects. Contact raises the stakes because the other object pushes back.

Autonomy is not a ladder from simple machine to synthetic person. It is an allocation of decisions. A rover may choose a safe route towards a destination chosen on Earth. A warehouse robot may navigate alone while a fleet system assigns its work. A surgical system may filter tremor while a surgeon remains in command. The useful question is which decisions have been delegated, within what operating domain, and how the system recovers when its assumptions fail.

The history follows one recurring trade. Early industrial robots became reliable by removing uncertainty from their surroundings. Better sensing, computation and control now let robots work in less structured places. Yet every step out of the cage adds variation, contact, human behaviour and safety obligations. The capability that expands robotics also creates its hardest remaining problems.

Reliability, safety and cost decide whether the bargain holds. The robot must keep working after dust, wear, moved furniture and imperfect parts enter the scene. It must fail without turning uncertainty into injury. It must save more effort, time or risk than integration and recovery consume. A spectacular motion is evidence of possibility. A maintained operation is evidence of robotics. The achievement is not freedom from constraints, but a design in which every important constraint has somewhere to go.

That is the book.

Why You Should Care

A robot arm can repeat a weld thousands of times and still fail because a panel arrived two centimetres away from the expected position. The stored motion may be flawless. The result is a perfect movement into empty space. Add a camera and the machine can measure the panel. Add calibration and it can relate pixels to millimetres. Add state estimation and it can decide where the panel probably is. Add feedback and it can correct its approach. Add force sensing and it can detect contact. Add recovery logic and it can stop asking an engineer to rescue every imperfect cycle.

That progression explains the field. Robotics joins mechanical engineering, electronics, control, geometry, software and, increasingly, machine learning, but it belongs to none of them alone. A fine planner cannot remove gearbox backlash. A precise camera is useless if its coordinate frame is wrong. A powerful motor makes a bad controller more dangerous. A clever hand may still be unable to lift the tool it has grasped. Physical intelligence must survive mass, friction, heat, delay and collision.

The scale is already industrial. The International Federation of Robotics recorded 542,000 new industrial robot installations in 2024 and 4.664 million industrial robots in operational use. Asia accounted for nearly three quarters of the new installations. Amazon said in 2025 that its operations had deployed a million robots. The dominant story is not a humanoid walking into a kitchen. It is specialised machinery becoming ordinary infrastructure.

Robotics also gives a better way to think about autonomy. Earth cannot joystick a Mars rover around each rock because communication is too slow. Human teams choose scientific goals and broad routes; Perseverance can build local terrain models, identify hazards and plan around them. In 2026 it also demonstrated onboard global localisation by matching panoramic camera images to orbital maps. The machine gained a new decision, not independence from the mission.

The same distinction applies closer to home. A warehouse robot may navigate without a driver while lacking authority to choose which customer order matters. A surgical system may translate a surgeon's movements with small, stable motions while remaining teleoperated. A collaborative arm may stop when contact exceeds a limit while having little high-level understanding. Robots contain different allocations of perception, motion and decision. Calling all of them smart hides the engineering.

The subject is also a clean way to separate a capability from a working system. A vision model may identify an object in a test set and fail when the gripper blocks the camera. A planner may find a collision-free path and ignore that the wheels cannot produce enough traction. A prototype may complete a task once; a product must complete it repeatedly, be maintained, recover from ordinary faults and justify its cost. Integration is not the clerical stage after invention. It is where invention becomes useful.

That makes robotics valuable when thinking about work. Occupations are bundles of tasks. Some are repetitive, some depend on dexterity, some require social judgement and some exist because exceptional cases occur. Robots tend to take task fragments first. Whether the remaining human job becomes safer, richer, more monotonous or merely faster depends on the surrounding workflow. Seeing the task bundle also prevents a successful demonstration of one movement from being mistaken for the automation of an occupation.

By the end of this book, a robot should no longer look like a character category. It should look like a designed agreement among the job, the environment, the body, the sensors, the controller and the people responsible for failure. That view is less dramatic than asking whether the machine is alive. It is far more useful, and the real machines become more impressive once you can see where the difficulty lives.

The Core Ideas

The Task Comes Before the Robot

The most important robotics decision is often made before anyone chooses a robot. It is the decision about the task: what must happen, how often, how quickly, under which conditions, with what tolerance and with what consequence when it goes wrong.

Consider moving a metal casting from a press. The object emerges hot, the motion repeats and the surroundings can be fenced. That combination helped make the first industrial robots commercially sensible. Now consider clearing a family kitchen after dinner. The objects vary, people walk through, food makes surfaces slippery, cupboards differ, fragile items sit beside rubbish and success includes knowing what should not be moved. Both tasks can be described as pick and place. Their engineering difficulty is barely related.

A task becomes tractable through its operating domain, the set of conditions in which the system is expected to work. Industrial automation often narrows that domain deliberately. Parts arrive within a known size range. Fixtures establish orientation. Lighting is controlled. Guards limit access. The robot does not overcome the whole physical world; it receives a smaller world with rules.

This is why the boundary between robot and environment is misleading. A bowl feeder that turns components into one orientation, a jig that guides a peg, a barcode that identifies a rack and a painted lane that simplifies navigation all perform work that could otherwise demand more sensing or reasoning. Intelligence can live in steel, tape and workflow. A cheap passive guide can beat an elaborate perception system because it removes uncertainty before the controller sees it.

The task definition also decides what should remain human. A hospital delivery robot may move medicines through corridors while staff load secure drawers, choose destinations and handle blocked lifts. A farm robot may identify weeds while a person confirms unusual cases. Human supervision is not always evidence that autonomy failed. It can be the correct allocation of rare judgement to people and repetitive motion to machines.

The mistake is to start with a fashionable body and search for a job. A humanoid, drone or arm can be an excellent answer where its geometry fits the task. It is an expensive question where it does not. Successful projects work backwards from required throughput, variation, access, safety, maintenance and economics. They compare the robot with conveyors, special-purpose machinery, better tooling, process redesign and human labour.

Core robotics therefore begins with a contract. The machine promises a specified performance inside a specified domain. The designer promises to control the parts of the world the machine cannot handle. Every later claim about perception, autonomy or intelligence is meaningful only inside that agreement. A robot without a declared domain is a promise whose difficult conditions have been left out.

Volume and variation usually pull the contract in opposite directions. High volume can justify expensive fixtures because the same certainty is bought millions of times. High variation rewards flexible sensing and tooling, but every added possibility must be tested and recovered. Low-volume, high-variation work often remains human because people arrive with a general body and can absorb unusual cases without a new integration project. The commercial frontier moves when hardware, software or process design changes that calculation, not when a robot merely performs the motion once.

A Body Is a Mechanical Decision

A robot's body is not packaging for its software. It determines which movements are possible, which forces can be applied, which surfaces can be crossed and which failures are likely.

A fixed six-axis arm trades mobility for stiffness, speed and a stable reference frame. A wheeled base moves efficiently across a warehouse but struggles with stairs and loose ground. Tracks spread load and cross rough surfaces at the price of friction and wear. Legs can step over obstacles, yet balancing consumes sensing, computation and energy before the machine has done any useful work. Drones gain access by continuously spending power to stay aloft. There is no neutral body.

Degrees of freedom describe the independent motions a mechanism can make. A rigid object in open three-dimensional space has six: three translations and three rotations. A manipulator needs enough joint freedom to place and orient its tool, but more joints create redundancy, self-collision possibilities and a larger space for planning. Extra capability carries an engineering bill.

The end effector often matters more than the arm. A suction cup can lift smooth boxes quickly and fail on porous cloth. Parallel fingers handle a wider range but demand better object positioning. A magnetic gripper is superb until the part is aluminium. Welding torches, screwdrivers, paint guns, cameras and surgical instruments turn the same arm into different machines. Tool changing can widen capability, while adding mass, interfaces, calibration and failure points.

Compliance changes how the body meets the world. A stiff mechanism can hold position accurately but transmit large forces during unexpected contact. Springs, flexible materials or controlled yielding let a robot absorb misalignment and interact more safely. Soft grippers wrap around fruit without requiring an exact model of every surface. The price may be lower force, slower motion or harder position control. Once again, the useful body depends on the job.

Scale matters. A tiny surgical mechanism can enter spaces a human hand cannot reach but has little room for motors and sensors. A mining robot can carry heavy tools but creates greater hazards and may need hydraulic power. Changing size alters inertia, heat flow, stiffness and power needs. A shape that works on a table does not become a field robot by being enlarged.

Humanoids make the trade unusually visible. Human buildings, tools and workstations provide a strong reason to copy human reach and locomotion. Yet the human body is an evolutionary compromise, not a factory specification. Two legs are difficult to stabilise. Hands are mechanically dense. Carrying actuators through moving limbs costs energy. A humanoid may avoid rebuilding the workplace while paying more for the robot.

Morphology is therefore part of the computation. A wheel solves balance mechanically. A funnel solves alignment geometrically. A compliant finger solves some uncertainty by deforming. Good robotics does not ask software to repair every bad mechanical choice.

Serial arms, parallel mechanisms and continuum bodies make the same point in different ways. A serial arm gains a broad workspace and accumulates link error. A parallel robot gains speed and stiffness through several supporting chains while accepting a smaller, more constrained workspace. A snake-like or soft body can enter narrow spaces while making precise shape estimation harder. There is no best robot body in isolation. There are bodies whose constraints line up well with a chosen environment and bodies that spend their whole control budget fighting their own form.

Sensors Measure, Robots Estimate

A camera does not tell a robot that a red cup is twenty centimetres to the left. It produces an array of light measurements. A lidar does not hand over a map. It returns ranges along beams, with missed reflections and noise. An encoder reports joint rotation. An inertial sensor reports acceleration and angular motion while carrying bias and drift. A force sensor reports deformation that must be converted into an estimate of load.

The difference between measurement and state is fundamental. State means the variables needed for action: where the robot is, how fast it is moving, where an object lies, whether a person is approaching, whether a grasp is secure. Some state can be measured closely. Much of it must be inferred from incomplete evidence.

Calibration makes the evidence agree. A camera mounted above a conveyor has a coordinate frame. The robot base has another. The tool tip has another. The object may be described in a world frame. The transformations among them must be measured. A small angular error at the camera can become a large position error at the end of a long reach. Many failures that appear intelligent are geometric bookkeeping failures.

Sensors also fail differently. Cameras provide rich appearance but depend on lighting and can lose depth on plain surfaces. Lidar measures shape and distance but may struggle with some reflective or transparent materials. Wheel encoders work well until the wheels slip. Inertial sensors respond quickly but drift. Force sensing detects contact after contact has begun. Combining them can produce a better estimate than trusting one alone.

Sensor fusion is not a vote. The estimator weighs evidence according to uncertainty and uses a model of how the system moves. A mobile robot can combine wheel odometry, inertial data, cameras and lidar. Odometry supplies smooth short-term motion and accumulates error. External features correct that drift when they can be recognised. The result remains a belief with uncertainty, not a perfect location.

Localisation and mapping reveal the circular problem. To build a map, the robot must know where it observed each feature. To know where it is, it often needs the map. Simultaneous localisation and mapping, or SLAM, estimates both together. The mathematics varies, but the engineering question stays: how can the system maintain a useful account of place while every input is imperfect?

Time is another sensor property. A measurement describes the world at an instant. Cameras expose, networks transmit and processors calculate. A fast robot acting on stale data may correct for a scene that no longer exists. Real-time design therefore cares about worst-case delay and synchronised clocks, not merely average computing speed.

Perception becomes harder around people because human motion carries intention. A person standing beside a corridor may stay still, turn or step into the path. The robot can estimate probabilities, maintain distance and choose conservative behaviour. It cannot remove uncertainty. The honest design goal is not certainty but bounded error, visible confidence and safe behaviour when confidence falls.

Sometimes the best next action is chosen for information rather than progress. A mobile robot can move to see around an occlusion. An arm can rotate an object before grasping it. A camera can change exposure or viewpoint. This is active perception: motion is used to improve the estimate that later motion will depend on. It exposes a deeper feature of embodiment. The robot does not passively receive a complete world and then act. Its actions determine which evidence becomes available, so sensing and motion form a loop before the tool touches anything.

Actuators Turn Decisions into Consequences

Robotics becomes physical at the actuator. An actuator converts energy into motion or force, and in doing so introduces limits that software cannot negotiate away.

Electric motors dominate many modern robots because they are controllable, efficient across useful ranges and easy to combine with encoders and power electronics. A motor produces torque and speed according to its design and operating point. Gearboxes trade speed for torque, while adding friction, backlash and compliance. Direct-drive joints remove much of the transmission and can improve response, but demand larger motors and careful thermal design.

Hydraulics move fluid under pressure and can produce high force in compact actuators. They suit heavy machinery and systems where power density matters. Leaks, pumps, valves, noise and maintenance come with the strength. Pneumatics use compressed air, making simple cylinders and grippers cheap and quick. Air's compressibility complicates fine control, though that same softness can be useful in contact.

Every actuator has an envelope. Continuous torque differs from peak torque. A motor may deliver a burst and overheat if asked to sustain it. A battery may hold enough energy for an hour while failing to supply a brief power demand. A joint may move quickly unloaded and slow under payload. Published numbers describe tests, not every combination of speed, posture and temperature.

Transmissions change what the controller feels. Backlash creates a dead zone when direction reverses. Friction can prevent small commands from moving the joint, then release suddenly. Flexible belts and long links store energy. Cable drives save distal mass and complicate tension. The mechanism has its own behaviour, which the controller must model or tolerate.

Contact makes actuation dangerous and useful. A free-moving arm mainly moves itself. At contact it can push, cut, lift, crush, polish or insert. Position control works well when the world is where the model says it is. If a rigid controller commands a tool through a surface that sits a few millimetres too high, force can rise sharply. Force control regulates interaction load. Impedance control makes the robot respond as though it had a chosen mechanical relationship between displacement and force, allowing it to yield rather than fight every mismatch.

The gripper closes the chain. Picking an object depends on friction, geometry, surface condition, centre of mass and acceleration after the lift. Humans adjust grip force through dense touch and compliance. A robot may use finger force, motor current, tactile arrays or vision, yet general grasping remains hard because the object moves when touched and often becomes hidden by the hand.

Energy limits autonomy in a literal sense. A mobile machine that works for twenty minutes and charges for two hours may be technically capable and operationally poor. Added batteries increase endurance and mass. Added mass increases required force and energy. Thermal management, charging, cable routing and wear decide how long the robot can keep its promises.

An actuator is where an abstract error becomes a bent part or injured person. That is why power, speed and payload must be considered with control and safety, never as isolated signs of capability.

Holding still can require effort too. A vertical arm may need continuous torque against gravity unless a brake, counterbalance or spring carries the load. A mobile robot may need braking authority on a slope. A flying robot must generate lift every second it remains airborne. These quiet demands shape battery life and fault behaviour. Loss of power is not one event across robotics: a wheeled base may coast to a stop, a suspended load may fall and a hydraulic joint may retain dangerous stored pressure. Safe design begins with what the body does when commands disappear.

Geometry Turns Intention into Motion

Suppose a robot must place a tool at a point with a specified orientation. The command describes the tool in space. The motors act at joints. Kinematics connects the two.

Forward kinematics asks where the tool will be if the joints take given positions. For a serial arm, each link transforms the next, so the final pose is built through a chain of rotations and translations. The calculation is direct once the geometry is known. Inverse kinematics asks the harder question: which joint positions will create the desired tool pose?

There may be several answers. An elbow can point up or down. A redundant arm can reach the same point through many postures. Some answers violate joint limits, collide with the robot or leave no room for the next move. Some place the mechanism near a singularity, a configuration where it loses an instantaneous direction of motion or requires extreme joint speeds to produce a modest tool velocity.

The Jacobian describes how small joint motions relate to tool motion and how forces at the tool relate to joint torques. It is one of the field's most useful bridges because it connects geometry, velocity, force and singularity. A general reader does not need to calculate one to understand the consequence: the same tool movement can be easy in one posture and difficult in another.

Mobile robots have geometry too. A car-like base cannot slide sideways. A differential-drive platform can turn in place but still obeys wheel constraints. A drone moves in three-dimensional space while its thrust direction depends on attitude. Planning must respect what the body can do, not merely whether a line fits between obstacles.

Configuration space provides the clean mental model. Instead of moving the whole shape through ordinary space, imagine each possible joint arrangement as a point in a higher-dimensional space. Obstacles become forbidden regions. Planning becomes a search for a continuous route through the allowed region. A six-joint arm begins with six dimensions before a gripper or mobile base is added, which explains why exhaustive search becomes expensive.

A path says where to go. A trajectory adds time, specifying positions, velocities and perhaps accelerations. The difference matters because the same path can be safe when slow and impossible when fast. Dynamics brings mass and inertia into the account. Accelerating a link changes the torques required at several joints; carrying a payload changes them again. Flexible structures vibrate. Gravity helps in one direction and opposes in another.

Engineers therefore move between geometric and physical descriptions. Kinematics filters impossible poses. Planning avoids collision. Trajectory generation respects speed and acceleration. Dynamics and control determine whether the real machine can execute the motion. A beautiful path in a simulation is a proposal until the motors, structure and contact conditions agree.

Redundancy can be used rather than merely tolerated. If an arm has more joint freedom than the tool task requires, it may keep the tool fixed while moving an elbow away from a person, avoiding a joint limit or improving leverage. The unused freedom becomes room for secondary objectives. That capacity is valuable and difficult because objectives can conflict. A posture that avoids collision may raise torque; one that improves dexterity may shorten the next available move. Geometry is where apparently simple instructions acquire consequences across the whole body.

Planning Chooses; Control Corrects

Planning and control are often blurred because both influence movement. They solve different time scales of the problem.

A planner chooses a route, action sequence or grasp from alternatives. For a mobile robot it may select a path through free space. For an arm it may choose a collision-free route through configuration space. For a warehouse fleet it may schedule corridors and priorities so individually valid routes do not produce collective deadlock. Planning uses models and predictions to decide what should happen.

Control deals with what is happening now. A position controller compares desired and measured joint position, calculates an error and adjusts motor commands. Proportional action responds to the current error. Integral action accumulates persistent error. Derivative action responds to how quickly error changes. The familiar PID pattern is useful because it turns error into correction, though real robot controllers often add model-based compensation, feedforward terms and limits.

Open-loop commands assume the machine and world will behave as predicted. Closed-loop control measures the result and corrects. Feedback can reject disturbances and model error, but it is not magic. Too much gain can make a system oscillate. Delay can destabilise a fast loop. Sensor noise can drive unnecessary motion. Saturated actuators cannot obey larger commands. The controller works inside the mechanical and timing limits of the system.

Higher layers close slower loops. A trajectory controller corrects motion over milliseconds. A navigation system may replan over seconds when an aisle closes. A fleet manager reallocates work over minutes. A maintenance system notices rising motor current over days. Autonomy is built from nested loops rather than one central mind.

This also explains why autonomy cannot be placed on a single scale. A drone may stabilise itself, hold position and avoid obstacles while a human chooses the route and camera target. A surgical system may control tremor and motion scaling while every purposeful movement comes from the surgeon. A rover may choose local paths while scientists choose destinations. The question is not how autonomous is it? The useful questions are: autonomous over which decision, for how long, under which conditions, with which override and with what recovery?

Modern learning methods can improve perception, policy selection or control. They may allow behaviour to be learned from data instead of written entirely as rules. Yet the learned component still enters a physical system with latency, force limits, safety constraints and an operating domain. A strong model in one layer does not dissolve bottlenecks elsewhere. Better language instruction does not increase wrist torque. Better object recognition does not guarantee a stable grasp.

Recovery is the neglected half of autonomy. A robot that performs ninety-nine cycles and needs an engineer on the hundredth may be a poor product. Useful autonomy detects when confidence has fallen, stops safely, retries where appropriate, requests the right kind of help and returns to service cheaply. Reliability is measured across failures, not around them.

Reactive and deliberative behaviour must also coexist. A planner can consider a route over seconds; obstacle avoidance may need a response in milliseconds. Safety functions may have to override both. The architecture therefore contains priorities and authority boundaries: which layer may command motion, which may veto it and what happens when components disagree. Autonomy fails when these boundaries are implicit. A system can possess good local behaviours and still behave badly because no layer has a coherent responsibility for the complete task.

Leaving the Cage Changes the Problem

Industrial robots became successful by receiving a disciplined world. They stood on fixed bases, followed repeated motions, handled known parts and worked behind guards. This was not primitive robotics. It was good systems engineering: uncertainty was reduced where reducing it was cheaper than perceiving it.

The pattern created the field's first scale. A programmable manipulator could handle hot castings, weld bodies, paint panels and tend machines because factories offered high repetition and a clear return on consistency. The International Federation of Robotics now counts millions of industrial robots in operation. Their importance lies in being dependable capital equipment, not in resembling people.

Better sensors, processors and algorithms have moved robots into warehouses, hospitals, farms, roads, homes, disaster sites and space. A mobile warehouse robot can localise and route around traffic. A cleaning robot can map rooms and return to charge. A rover can evaluate local terrain without waiting for Earth. These systems handle more uncertainty on board, which widens the operating domain.

But the world outside the cage brings categories of difficulty that a better planner alone cannot settle. Surfaces change friction. Weather affects sensing. Objects deform. People behave unpredictably. Doors close. Lifts fail. Wireless networks drop. A hospital corridor is also a social space in which blocking the way can matter more than reaching the destination quickly.

Human-robot interaction therefore becomes part of control. The machine must signal intent, maintain appropriate distance, accept interruption and avoid forcing people to learn an obscure private language. Trust must be calibrated. A robot that looks confident while uncertain encourages misuse. One that stops at every harmless ambiguity becomes intolerable. The design problem includes what the human believes the robot can do.

Safety expands from the robot to the application. A small arm with force limits can still carry a sharp tool. A mobile base can trap a foot. A safe joint can move an unsafe workpiece. Current industrial standards separate requirements for the robot from requirements for the integrated application because hazards emerge from the complete cell, task and foreseeable misuse. The label collaborative describes a designed interaction, not an amulet.

Organisation matters too. A warehouse fleet requires traffic rules, charging, maintenance, software updates and stations designed around people. A hospital robot needs lift access, cleaning procedures, secure payloads and someone responsible for exceptions. A machine may pass its laboratory tests and fail as a service because nobody owns recovery at three in the morning.

This repays the opening condition. Robotics first prospered by deciding what the world would do for the robot. Modern autonomy asks the robot to do more for itself. Every new freedom therefore transfers work from fixtures, guards and operators into sensing, estimation, planning, interaction and assurance. Leaving the cage is progress. It is also a change in who carries the uncertainty.

The deployment statistics reflect the old and new worlds unevenly. Industrial robots can be counted under stable categories because their forms and applications are mature. Service robotics is broader, and IFR warns that its figures come from changing supplier samples rather than a census of the whole industry. The asymmetry is informative. The best-measured part of robotics is the part with the clearest contracts. The most expansive claims often concern machines whose operating domains, intervention rates and business models are still being negotiated.

How It Actually Works

Hot metal, stored motions

In 1961 a Unimate began work at a General Motors plant in New Jersey. The machine was a large hydraulic arm controlled by step-by-step commands stored on a magnetic drum. It handled hot die-cast metal, a task with heat, repetition and a clear path. Those conditions explain more than the date. The first successful industrial robot did not enter a normal workplace and adapt to whatever it found. The workplace supplied a repeatable object, a repeatable sequence and physical separation from people.

That arrangement became the industrial cell. The robot sits on a known base. Fixtures locate workpieces. Interlocks confirm that guards are closed. Tooling performs a narrow operation. A controller executes a programme with defined speeds, positions and signals. The cell may include conveyors, clamps, cameras, feeders and process equipment. Calling the arm the robot is convenient; the productive machine is the integrated cell.

The operating cycle begins with conditions. Is the correct part present? Is the tool attached? Is air pressure available? Is the guarded space clear? A programmable logic controller may coordinate these events while the robot controller handles motion. Once the prerequisites are satisfied, the arm moves through taught points or generated trajectories, operates the tool, checks completion and hands control to the next stage.

Early systems were blind by modern standards, but the factory gave them predictability. That strategy remains powerful. A high-volume welding line can justify special fixtures and long integration because the same cycle will run for years. A low-volume workshop with changing parts may favour vision, flexible tooling or a person. The operating domain and economics choose the architecture.

Research laboratories then attacked the uncertainty that factories had removed. SRI's Shakey project, running from 1966 to 1972, joined sensing, route planning and mobile action in a simplified environment. Stanford's electrically powered computer-controlled arm, developed in 1969, later assembled a Model T water pump using optical and contact sensing. Neither machine was a general worker. Together they made two durable questions visible: can a robot form a usable account of the world, and can it turn that account into controlled physical action? Industrial robotics and autonomous robotics grew by answering those questions under different contracts.

Teaching a mechanism where the world is

A robot cannot act usefully until its internal coordinates correspond to physical space. Integrators establish a base frame, tool frame and frames for fixtures or workpieces. They measure the tool centre point, the effective working point of a gripper, torch or screwdriver. They calibrate cameras and check how payload bends the mechanism. These steps are less visible than programming and often decide accuracy.

Industrial arms can be taught by moving them to positions and recording the joint values. A programmer may then connect those points with joint motions, straight tool paths or arcs, adding speed, acceleration, tool commands and conditions. Offline programming uses a digital model of the cell so much of the work can be prepared before production stops. In either case, the stored path assumes that the real cell matches the model.

The distinction between repeatability and accuracy appears here. An arm can return consistently to a biased point. Calibration can compensate if the bias is stable. A new tool, changed temperature, worn gearbox or shifted fixture can move the error. The correct performance measure belongs to the complete task: did the weld land within tolerance, did the insertion succeed, did the seal remain unbroken?

Sensors widen the cell's tolerance. A camera can locate a part that arrives within a region rather than at one exact point. A force sensor can detect when a peg meets the edge of a hole. A laser tracker can inspect a surface. The robot becomes less dependent on perfect presentation, though every sensor adds calibration, timing and failure modes.

From light and rotation to state

A mobile robot cannot rely on a fixed base. It must continually estimate where it is and what surrounds it. Wheel encoders estimate motion from rotation. An inertial measurement unit tracks acceleration and angular rate. Cameras observe texture and features. Lidar returns ranges to surfaces. Satellite navigation may provide global position outdoors. Each source contributes evidence and carries characteristic error.

Odometry integrates small motions to estimate a changing pose. The estimate drifts because wheels slip, diameters differ, surfaces deform and sensor errors accumulate. A robot can correct by recognising landmarks in a map, matching a new scan to an earlier scan or observing an external reference. The estimator must also maintain uncertainty. A point estimate without confidence can make a lost robot look certain.

SLAM is used when the map and pose are both unknown. The robot moves, observes features and estimates a trajectory and map that best explain the measurements. Revisiting a known place creates a loop closure, which can correct drift across the whole route. False matches can damage the map, so systems use geometric checks and multiple cues rather than accepting every resemblance.

Perseverance gives the problem at planetary scale. Its visual odometry follows changes in terrain features while accounting for wheel slip, but small errors accumulate over long drives. In February 2026 the rover used Mars Global Localization to capture a panorama and match it against onboard orbital imagery. The calculation took about two minutes and located the rover within roughly twenty-five centimetres. The new method did not replace local navigation. It provided a way to reset accumulated global position uncertainty.

Choosing a feasible movement

Once the robot has an estimate of state, it needs a movement that its body can execute. An arm first solves for joint configurations that place the tool where required. A mobile base searches for a route compatible with its turning constraints. A legged robot chooses footholds while maintaining balance. A drone chooses attitude and thrust that move it without exhausting its margin against wind.

Collision checking must include the whole body and payload, not merely the tool point. An arm can place its gripper in free space while an elbow strikes a guard. A mobile robot can clear an obstacle and swing a carried load into it. Planners therefore search configuration space, where each point describes a full body arrangement.

High-dimensional search is difficult, so planning methods exploit structure. Sampling-based methods test selected configurations and connect feasible regions rather than fill the entire space with a grid. Optimisation methods begin with a candidate trajectory and reduce collision, smoothness or energy costs under constraints. Grid and graph searches remain useful for many mobile maps. No one planner is best for every machine, environment and deadline.

The output must become a timed trajectory. Speed and acceleration limits protect motors and loads. Jerk, the rate of change of acceleration, matters for fragile payloads and human comfort. Dynamic limits matter when motion is fast or the body is heavy. A feasible geometric path may still demand more torque, grip or traction than the machine can provide.

Closing the fast loop

The trajectory is a reference, not a guarantee. Motors receive commands; the structure flexes; friction changes; external forces arrive. Feedback control compares desired and measured motion and adjusts the command.

Joint controllers run quickly because delayed correction becomes poor correction. They may combine model-based feedforward, which predicts the effort required, with feedback that removes residual error. Limits constrain speed, torque, current and temperature. Watchdogs detect missing messages or stalled computation. Emergency functions bring the system towards a safe state when ordinary control can no longer be trusted.

Contact requires another mode of behaviour. During free-space motion, position is the main objective. During polishing, insertion or human assistance, interaction force matters. A controller may regulate force along one direction and position along another. Impedance control makes the relationship between motion and force resemble a chosen spring and damper. The robot can then meet a slightly misplaced surface without trying to drive through it rigidly.

Compliant hardware can share the burden. Elastic elements absorb shocks and provide force information through their deflection. Soft fingers conform to uncertain shapes. Rounded surfaces reduce some contact hazards. These features do not remove the need for control, but they change the consequences of imperfect estimates.

A grasp is a small control system. The gripper approaches, contacts, closes to a force or position, checks whether the object is present and monitors it during motion. A failed pick may trigger a retry from another angle, send the item to an exception station or request a person. The recovery policy is part of the cycle time.

Making many components behave as one machine

Modern robots contain distributed software. Device drivers communicate with motors and sensors. Estimators calculate pose. perception components identify objects or free space. Planners choose routes. Controllers send commands. User interfaces, safety systems, logging and fleet services surround them. Middleware such as ROS 2 provides patterns for modular nodes to exchange continuous data, request services and run longer actions.

Modularity speeds development and creates interfaces. One component may use metres while another expects millimetres. A camera frame may be labelled incorrectly. Clocks may disagree. A high-bandwidth sensor may overload a network. A planner may update more slowly than the controller assumes. Robust systems define units, frames, timing and failure behaviour as carefully as algorithms.

Simulation helps teams test before risking hardware. A digital model can exercise kinematics, collision checking, traffic rules and fault responses. Automated tests can run thousands of cases. Synthetic scenes can expose rare conditions. Hardware-in-the-loop tests connect real controllers or components to simulated surroundings.

The simulation gap remains. Friction, cable drag, glare, dust, deformable objects and human improvisation are hard to model with enough fidelity. A system trained or tested only in simulation may have learned the simulator's conveniences. Teams therefore move through staged trials: simulation, controlled hardware tests, pilot deployments and monitored expansion of the operating domain.

Configuration management matters after release. A software update can alter braking distance or message timing. A replacement camera can change calibration. Logs must identify which code, maps and parameters produced an event. Robotics products need the discipline of software operations plus the discipline of machinery maintenance.

A fleet is a different robot

Warehouses reveal what changes when hundreds of mobile robots share a building. Each unit may localise, avoid immediate obstacles and follow assigned routes. A fleet manager allocates work, reserves narrow passages, balances charging and prevents deadlock. The effective system includes racks, stations, wireless networks, floor design, people and exception processes.

Amazon's million-robot milestone illustrates scale but not a universal architecture. Such deployments work because the building and workflow are designed around the fleet. Inventory is presented in compatible storage. Human stations receive work in planned sequences. Maintenance teams recover failed units. Software can optimise traffic across the network. The unit of performance is orders moved through the operation, not the elegance of one robot's path.

A fleet also changes supervision. One person may monitor many machines and intervene only on exceptional cases. The economically important measure is often interventions per operating hour and the expertise needed to resolve them. A system that appears highly autonomous but creates frequent, difficult exceptions may consume more labour than it saves.

Charging becomes scheduling. Sending every robot to charge at the same battery threshold can create a queue and reduce throughput. Batteries age differently. Busy zones create traffic. Fleet management uses predicted workload, state of charge and station availability to maintain capacity. The intelligence is organisational as much as onboard.

Working around people

Traditional industrial cells use separation because a fast, heavy arm can generate dangerous force. Human-robot collaboration allows closer work through defined protective measures. A robot may stop when a person enters a monitored space, reduce speed as separation shrinks, limit power and force, or allow hand-guided operation. The choice depends on the application.

Safety analysis begins with hazards. The robot can strike, crush, trap or eject. The tool can cut or burn. A payload can fall. Unexpected restart, stored energy and maintenance access create risks beyond normal motion. Integrators assess severity and exposure, apply protective measures, validate them and provide information for use. Current ISO 10218 standards separate the industrial robot as partly completed machinery from the completed robot application and cell. That division reflects reality: the same arm can be safe in one integration and hazardous in another.

Interaction also has a social layer. A hospital delivery robot must indicate direction, yield appropriately and avoid blocking urgent movement. A person needs to understand whether the machine has seen them and what it will do next. Sound, light, motion and screen messages can communicate state, but signals must be consistent and readable under pressure.

Human factors extend to work design. If a robot sets an unforgiving pace, moves awkward tasks onto people or hides the reason for stoppages, technical success can produce poor work. Operators often possess practical knowledge about variation and failure that designers lack. Bringing them into deployment can improve both safety and throughput.

Maintenance, recovery and the cost of the last per cent

A deployed robot spends part of its life dirty, worn, blocked or confused. Lenses collect dust. Wheels lose tread. Gearboxes develop play. Grippers wear. Maps become stale. Networks drop. Pallets arrive outside marked zones. Someone leaves a ladder where no ladder belongs.

Good systems detect degradation before it becomes damage. Motor current, vibration, temperature and cycle-time trends can reveal wear. Self-checks test sensors and protective devices. Preventive maintenance replaces parts on schedule; condition-based maintenance uses evidence from the machine. Spare components and trained technicians determine how quickly work resumes.

Recovery should be designed like normal operation. The robot needs safe stop states, accessible release procedures, clear fault codes and tools that let a trained person diagnose events. Remote support can help, but a network loss must not create an unsafe dependency. Some systems fall back to reduced speed or limited capability when a sensor is unavailable. Others must stop because continued operation cannot be assured.

Economics absorbs all these details. The purchase price is only one line. Tooling, integration, guarding, validation, training, software, building changes, energy, maintenance, downtime and process redesign can dominate. Benefits include throughput, quality, safety, traceability, space and labour allocation. The comparison is against the best alternative process, not against doing nothing.

This is why the last few percentage points of reliability can decide the whole project. A laboratory system may prove that a task can be done. A product proves that it can be done across shifts, operators, wear and ordinary disorder. Physical reliability is not the quiet phase after innovation. It is the threshold between a demonstration and infrastructure.

Proving the operating domain

Before release, teams must turn the task contract into tests. Nominal trials show the intended cycle. Boundary trials probe the edges: the darkest allowed lighting, the heaviest payload, the smallest obstacle, the longest communication delay, the least favourable battery state and the most awkward permitted posture. Fault injection checks what happens when a sensor freezes, a message arrives late or an actuator reports an impossible state.

The test set must include combinations. A camera may work in dim light and a gripper may work on a glossy object, yet the combination can fail because the camera loses the edge the grasp planner needs. A mobile robot may stop safely on a clean floor and localise reliably in a crowded corridor, yet fail when emergency braking causes wheel slip that corrupts its pose estimate. Physical systems produce interactions that component tests miss.

Safety validation asks whether protective measures achieve their claimed performance, including stopping distance, separation, force and access control. Operational validation asks whether ordinary staff can recognise faults, clear them safely and resume work. Cybersecurity, software updates and data handling enter where connected robots can affect physical behaviour or expose sensitive environments.

Deployment should widen in controlled steps. A pilot starts with limited routes, products or shifts. Logs reveal intervention patterns. Engineers remove recurring causes, not merely reset machines. The operating domain expands only when evidence supports it. This discipline can look cautious beside a polished demonstration. It is how a robot earns the right to become boring.

How we know

Robotics claims eventually meet hardware, which makes much of the field testable. Kinematics can be checked against measured pose. Controllers can be evaluated through tracking error, stability and disturbance response. Perception can be tested on recorded data and physical trials. Deployed systems produce logs of faults, interventions, cycle times and maintenance. Standards define safety requirements, while industry bodies compile installation statistics.

The evidence has gaps. Laboratory benchmarks often simplify lighting, objects, terrain and human behaviour. Videos select successful runs. Commercial intervention and failure data are commonly proprietary. Industrial robot statistics are comparatively standardised, while service robot figures are based on supplier samples whose composition changes. Broad claims about general-purpose autonomy therefore deserve more caution than claims about a fixed cell. Independent trials across different hardware, sites and operators are rarer than benchmark results and should carry greater weight when they exist. The strongest evidence is repeated physical performance across the full operating domain, including recovery from the cases that a demonstration would cut away.

What People Get Wrong

“A robot has to look like a person”

Fiction trained us to identify robots by body shape: head, torso, arms and perhaps a metallic voice. Engineering classifies them by what they can do. A fixed manipulator, an autonomous mobile platform, a milking system and a pipe crawler can all qualify without sharing a silhouette.

The human form is useful where the environment has been built for human reach, stairs, handles and tools. It is inefficient where the task permits a specialised body. Wheels use less energy than legs on a level floor. A delta mechanism sorts small products faster than a human-shaped arm. A suction tool can outperform a five-fingered hand on boxes.

The misconception matters because appearance distorts investment and judgement. A familiar face makes limited machinery seem general. An ungainly fixture can hide extraordinary performance. Evaluate the task, operating domain, payload, speed, supervision and recovery. Resemblance is one mechanical strategy, not the definition or destination of robotics.

The word robot entered public life through fictional artificial workers, so the image arrived before the engineering category. That history still pulls attention towards characters. The correction redirects attention to physical agency and delegated decision-making, which are present even when there is no face to look at.

“Automation and robotics are the same thing”

Automation is the wider category. A thermostat, bottling line, software script and railway signalling system can operate automatically without being robots. Robotics adds an actuated mechanism with a degree of autonomy for locomotion, manipulation or positioning. The boundary can be awkward, but it prevents every controlled machine from becoming a robot by marketing decree.

A conveyor moves material automatically along a fixed route. An automated guided vehicle may follow markers or external guidance. An autonomous mobile robot can choose routes using onboard sensing and navigation. All can solve the same logistics problem, and the simpler system may be better.

Confusing the terms encourages unnecessary complexity. A business may buy a mobile robot where a conveyor would be cheaper, or demand a robotic arm where a dedicated mechanism would be faster. The useful question is not whether a process can be robotised. It is which degree of programmability, motion and adaptation produces the best complete system.

Marketing blurs the line because robot sounds more advanced than automated equipment. Standards draw boundaries for statistics and safety, but projects should compare architectures rather than defend labels. A fixed machine that solves the task cleanly is not inferior because it lacks enough autonomy to enter a robotics census.

“Sensors tell the robot what is there”

Sensors return signals, not facts. A camera records light. Lidar measures returned pulses. Encoders measure motion. A force sensor measures deformation. The robot must calibrate those readings, align coordinate frames, account for delay and infer the state needed for action.

This gap explains many strange failures. A camera can identify the correct object while reporting its pose in the wrong frame. A mobile robot can have accurate wheel readings and drift because the wheels slipped. A transparent surface can confuse range sensing. A detector can see a person after network delay has made the position stale.

The correction matters whenever a percentage is used to describe perception. High classification accuracy does not establish that a robot can grasp safely, navigate reliably or know when it is uncertain. Ask what was measured, under which conditions, how state was estimated and what the machine does when its evidence conflicts.

Humans encourage the confusion because seeing feels immediate. Our own perception is also constructed, but years of embodied experience hide the process. Robots expose every calibration, frame transformation and confidence estimate, which makes ordinary visual competence look less like a camera feature and more like a continuing act of inference.

“A precise robot is an accurate robot”

Repeatability and accuracy answer different questions. Repeatability asks whether the machine returns to the same commanded condition. Accuracy asks how close that condition is to the intended truth. A robot can repeat a point with tiny variation and miss the target by the same two millimetres every time.

Industrial systems often value repeatability because stable bias can be calibrated out and fixtures can establish the task geometry. Yet performance changes with posture, payload, speed, temperature, wear and tool length. A catalogue number measured under a standard test cannot guarantee every application.

The distinction protects against buying by headline specification. The relevant quantity is the task's error budget: robot motion, tool calibration, part presentation, sensing, fixture tolerance and process variation combined. A more repeatable arm may lose to a better-designed cell. The product being judged is the completed operation, not one number on the mechanism.

The confusion persists because precise motion is easy to see while true position needs a reference. A robot tracing the same elegant path in the air looks impressive even when the entire path is shifted. Metrology supplies the external truth that repetition alone cannot provide.

“Artificial intelligence is the robot's brain”

The brain metaphor implies a single centre that perceives, thinks and commands a passive body. Real robots distribute decisions across mechanisms, controllers, safety functions, estimators, planners and human systems. A reflex may run in a motor drive. A force limit may be enforced independently of the main computer. A fleet manager may make a more important decision than any individual robot.

Machine learning can improve object detection, language instruction, grasp selection or control. Those gains remain bounded by calibration, latency, power, geometry and assurance. A language model cannot create traction. A learned policy cannot use torque the actuator does not possess. A strong perception model can still be mounted in the wrong place.

This correction preserves the boundary between robotics and AI. AI methods may be valuable components. Robotics is the larger physical system that decides whether those components can act repeatedly, safely and economically in the world.

The metaphor also hides mechanical intelligence. A compliant linkage, differential gear or shaped guide can produce adaptive behaviour without a learned model. Calling every useful response AI makes the software look cleverer and the engineering around it disappear. The whole system deserves the credit and the blame.

“Collaborative robots are safe by themselves”

A collaborative arm is often smaller, slower and designed with functions such as monitored stopping or power and force limiting. None of those properties makes every application safe. The tool can be sharp. The workpiece can be heavy. The layout can create a trapping point. A fault during maintenance can release stored energy.

Current industrial standards distinguish the robot from the integrated application. One part addresses the robot as machinery; another addresses the cell, end effector, commissioning, operation and maintenance. That separation exists because risk emerges from the combination.

The label collaborative should therefore prompt questions, not relaxation. Which collaborative method is being used? What speed and force were validated? What happens near the tool? Can a person enter unexpectedly? How is foreseeable misuse handled? Human proximity is an application-level engineering claim that must be demonstrated for the task.

The myth became persuasive because small rounded arms look benign beside older fenced machines. Appearance lowers vigilance. Risk assessment restores the missing objects: payload, edges, pinch points, speed, access and stored energy. Safety belongs to the interaction that can occur, not to the colour or size of the arm.

“More autonomy always means a better robot”

Autonomy can remove delay, reduce supervision and let a machine handle variation. It can also make behaviour harder to verify, exceptions harder to diagnose and responsibility less clear. The correct amount depends on the decision and operating domain.

A Mars rover benefits from local navigation because Earth cannot respond in real time. A surgical system may be better with a surgeon directing purposeful motion while the machine scales movement and filters tremor. A warehouse fleet may navigate autonomously while human staff handle damaged goods. Full independence would add cost without improving the service.

Intervention burden is a better measure than an autonomy label. How often does the machine need help? Can one person supervise many units? Does recovery need an expert? Does the system recognise uncertainty before failure? Useful autonomy moves human attention to the cases where judgement has the highest value. It is delegated competence, not a ceremonial march away from people.

The misconception survives because autonomy is marketed as a rank. In operation it is a design variable. Removing a human decision is progress only when the replacement is reliable, inspectable and cheaper than keeping the decision human. A lower-autonomy system can be the more advanced product when it divides responsibility well.

Use It

Find where the world has been engineered for the machine

When a robot looks uncannily competent, inspect its surroundings. Are objects presented in fixed trays? Is lighting controlled? Are floors mapped and kept clear? Are people excluded? Does a guide rail remove the need for fine perception? These conditions reveal how the system divides difficulty.

That division is not cheating. It is engineering. A passive funnel can align a component more reliably than a camera and planner. A rack designed for a mobile platform can save battery and sensing. A standard container can turn an awkward grasping problem into routine handling. The intelligent choice is often to change the world once rather than ask the robot to solve the same uncertainty on every cycle.

Use this lens on demonstrations. List the assumptions supplied for free, then imagine one being violated. Move the object, change the surface, add a person, reduce the light or interrupt the network. The response tells you whether you are watching a narrow but mature system or a broad but fragile capability.

Separate measurement from belief

When a machine claims to see, locate or understand, ask what it measured and what it inferred. A range reading, pixel value or wheel rotation is evidence. Object pose, free space and robot location are estimates constructed from evidence and models.

This distinction improves diagnosis. A robot may have functioning cameras and still fail because calibration is wrong. It may recognise an object and misjudge depth. It may localise well in a mapped corridor and become uncertain after furniture moves. It may fuse several sensors and give too much weight to one that has entered a bad condition.

Look for uncertainty as an output, not an embarrassment. A well-designed system knows when evidence is weak, slows down, seeks another view or requests help. Confidence that never changes is often a sign that the architecture has no honest representation of doubt.

Ask what happens at contact

Robotic videos are full of free-space motion because it is easy to make smooth and satisfying. Useful work often begins where the video becomes dangerous: the tool touches, the gripper closes, the wheel slips or the machine shares force with a person.

At contact, small position errors can create large forces. Friction decides whether a grasp holds. Compliance decides whether misalignment becomes a successful insertion or a jam. The shape of the tool can matter more than the intelligence of the planner. A machine that reaches an object has completed only the first half of manipulation.

So ask what the robot can feel, how it regulates force, what yields, what can be trapped and how energy is removed after a fault. This lens distinguishes motion theatre from physical competence. It also explains why a slower, softer robot may be more useful than a faster one around uncertain objects and people.

Look for the human outside the frame

A robot described as autonomous may rely on people who label maps, load fixtures, clear faults, approve actions, maintain batteries and handle exceptions. The human work is often outside the promotional shot rather than absent.

Map the whole service. Who defines the task? Who monitors the fleet? Who cleans the sensors? Who takes over when localisation fails? Who is responsible when the robot blocks a fire door? A machine that shifts labour from visible operators to hidden support has changed the job, not removed it.

This lens also guards against crude claims about employment. Automation acts on tasks and workflows. It can remove hazardous lifting, create maintenance work, intensify pace, narrow discretion or free time for judgement. The result depends on design choices and bargaining, not on the mere presence of a robot.

Judge autonomy by recovery

The polished run shows the nominal behaviour. Deployment is decided by what happens after a tote is crushed, a door remains closed, a person steps into the route or a gripper misses. Ask how the system detects the problem, enters a safe state and returns to work.

Intervention frequency is only one measure. Intervention difficulty matters too. A fault that any operator can clear in thirty seconds is different from one that requires remote engineering support. Recovery should preserve evidence through logs and clear fault states rather than erase the event with a reboot.

A strong autonomy claim therefore includes boundaries. It says which conditions the system handles, which conditions cause a controlled stop and what supervision remains. A machine that admits uncertainty before causing damage is more mature than one that completes more demonstrations and fails opaquely.

Follow the bottleneck, not the breakthrough

Robotics headlines often promote one component: a new hand, battery, vision model or walking controller. The complete system advances only when that improvement reaches the current limiting constraint. A better gripper does little if object presentation dominates failure. Longer battery life does little if the machine needs manual rescue every twenty minutes. Faster planning does little when safety rules cap speed.

Trace the dependency chain from task completion backwards. Which failure consumes the most downtime? Which assumption narrows the operating domain? Which cost prevents scale? The answer may be a cheap connector, a cleaning routine or a redesigned station rather than the component attracting attention.

This lens also improves forecasts. Progress propagates unevenly. A dramatic result in simulation may wait for sensors, actuators, standards and maintenance practices to catch up. A modest improvement in reliability may produce rapid adoption because every other layer was ready. The important curve is system performance across real conditions, not the isolated benchmark of the newest part. Adoption begins when the slowest layer becomes good enough.

The limits

Robotics does not provide a universal recipe for replacing physical work. Some environments remain too varied, some objects too deformable, some tasks too low-volume and some failures too costly. Human bodies combine mobility, touch, perception and adaptation at a price and energy efficiency that engineered systems still struggle to match.

The field also cannot settle social questions by engineering. A safe machine can be used in a harmful labour system. A productive robot can concentrate power or spread capability. A care robot can support independence or become an excuse to withdraw human contact. Technical performance does not decide which outcome is acceptable.

Nor does progress in one component guarantee progress in the complete system. Better batteries may extend runtime without improving grasping. Better language interfaces may simplify instruction without improving reliability. Better tactile sensing may generate more data than the controller can use. Bottlenecks move.

Evidence becomes thinner as claims become broader. A fixed cell can be tested across millions of cycles. A general-purpose system faces an open set of situations. No number of selected videos proves competence outside the demonstrated operating domain. The more universal the promise, the more varied and adversarial the test should be.

The one thing to keep

Keep the contract.

A robot is a physical agreement among a task, a body and a world. The task defines success and failure. The body chooses the available motion and force. Sensors provide uncertain evidence. Estimation constructs a working state. Planning chooses among possibilities. Control corrects the real machine. People and infrastructure carry everything the robot has not been designed to handle.

This changes how a new robot should be read. Do not begin with the face or the word intelligence. Ask what job is completed, under which conditions, with what tools, at what intervention rate and with what consequence when the model is wrong. Find where uncertainty has been removed, where it has been estimated and where it has been handed to a person.

Early industrial robotics made the bargain explicit by placing machines in cages and designing the work around them. Modern autonomy widens the bargain, allowing robots to carry more uncertainty themselves. It does not abolish the bargain. It moves its boundary.

Once you see that boundary, robotics stops looking like a contest to build an artificial person. It becomes the harder and more useful discipline of arranging bodies, decisions and environments so that physical work happens reliably. The robot is the visible part. The agreement is the machine.

Terms

Actuator. A device that converts energy into controlled motion or force. Electric motors, hydraulic cylinders and pneumatic cylinders are common forms. Actuation is what lets a robot change the world rather than only observe it.

Autonomy. The ability to perform intended tasks from current state and sensing without continuous human intervention. It applies to particular decisions and operating conditions, not to the machine as a claim of general independence.

Calibration. The process of relating measurements and coordinate frames to physical references. Calibration determines whether a camera, joint, tool and workpiece agree about where things are, and whether that agreement remains valid after change or wear.

Closed-loop control. Control that measures the result of action and uses the difference between desired and measured behaviour to adjust later commands. The loop corrects disturbance and model error within physical limits.

Collaborative application. A robot application designed for defined forms of human-robot collaboration using assessed protective measures. Collaboration describes the integrated task and cell, not an automatically safe property of an arm.

Compliance. The ability of a mechanism or controller to yield under force. Compliance can come from springs, flexible structures, soft materials or control laws that regulate the relationship between displacement and load.

Configuration space. A mathematical space in which each point represents one complete robot configuration. Obstacles become forbidden regions, allowing motion planning to search for a route for the whole body.

Degree of freedom. One independent coordinate of possible motion. Extra degrees of freedom can widen reach and redundancy while increasing mechanical complexity, planning burden and opportunities for self-collision.

Dynamics. The study of motion under forces and torques, including mass and inertia. Kinematics describes possible movement; dynamics describes the physical effort and response involved in producing it.

Encoder. A sensor that measures rotation or linear displacement. Joint and wheel encoders provide position and velocity evidence for feedback control and odometry.

End effector. The tool at the working end of a manipulator. Grippers, suction cups, welding torches, screwdrivers, cameras and surgical instruments create different robots from the same arm.

Force control. Control that regulates interaction force rather than position alone. It is used for pressing, polishing, insertion and other tasks in which contact load matters.

Forward kinematics. Calculating the pose of the tool or body from known joint positions. It follows the mechanism's geometric chain from base through links and joints.

Human-robot interaction. The study and design of how people and robots communicate, coordinate, share space and divide decisions. It includes physical safety, interfaces, trust, supervision and work organisation.

Impedance control. A method that controls the dynamic relationship between motion and force, making the robot respond like a chosen spring and damper. It supports compliant contact with uncertain surroundings.

Inverse kinematics. Calculating joint positions that place a tool at a desired pose. The problem can have several solutions, no solution or solutions that violate limits and collision constraints.

Jacobian. A matrix relating small joint motions to tool motion and tool forces to joint torques. It reveals directional capability and becomes ill-conditioned near singular configurations.

Lidar. A sensor that measures distance using emitted light and its return. Lidar can provide precise geometry for mapping and obstacle detection, while reflective or transparent surfaces can cause difficulty.

Localisation. Estimating a robot's pose within a map or reference frame. Localisation combines measurements and motion models and should preserve uncertainty rather than report false certainty.

Manipulator. A mechanism of linked joints designed to position and orient an end effector. Manipulators may be fixed, mounted on mobile bases or integrated into larger machines.

Mobile robot. A robot able to travel under its own control. Wheels, tracks, legs, flight and underwater propulsion create different mobility constraints, sensing needs and energy costs.

Morphology. The physical form and arrangement of a robot's body. Morphology determines access, stability, force, efficiency and which problems can be solved mechanically before software begins.

Odometry. Estimating change in pose from motion measurements, often wheel or visual data. Odometry is smooth and useful over short periods but accumulates error through noise, slip and bias.

Path planning. Finding a feasible route between configurations or locations while respecting obstacles and constraints. A path specifies geometry; a trajectory adds timing and dynamic limits.

Payload. The mass or load a robot is designed to carry or manipulate under stated conditions. Payload interacts with reach, speed, acceleration, posture and thermal limits, so the headline maximum rarely applies across the full workspace.

Repeatability. The consistency with which a robot returns to the same commanded condition. Repeatability differs from accuracy, which measures closeness to the intended true condition.

Sensor fusion. Combining several measurement sources and models into a state estimate. Fusion exploits different strengths and failure modes, while remaining dependent on calibration, timing and uncertainty assumptions.

Singularity. A posture where the mechanism's local mobility degenerates, making some tool directions unavailable or demanding excessive joint motion. Singularities change dexterity, force transmission, planning and control.

SLAM. Simultaneous localisation and mapping. A family of methods that estimates a map and the robot's pose within it from uncertain observations and motion.

Trajectory. A path with timing, describing how position, velocity and often acceleration should change. Controllers track trajectories subject to actuator, contact and safety constraints.

Go Deeper

Kevin M. Lynch and Frank C. Park, Modern Robotics: Mechanics, Planning, and Control (Cambridge University Press, 2017). Start here for the mathematical skeleton of the field. The book connects rigid-body motion, kinematics, dynamics, planning and control in one geometric framework, with exercises and supporting material. It is a university text and uses linear algebra, but its sequence makes clear why each tool is needed. Read the early chapters with a simple arm or simulation beside you and the abstractions become movements. The authors also provide open supporting software and videos, which make it the most approachable route from intuition to formal robotics. Work one example by hand before relying on the supplied code.

Sebastian Thrun, Wolfram Burgard and Dieter Fox, Probabilistic Robotics (MIT Press, 2005). Read this for the problem that sensors create: a robot never knows its state perfectly. The authors develop localisation, mapping, tracking and decision-making as inference under uncertainty. The examples predate the current learning boom, and the mathematics is substantial. The mental model remains indispensable because newer perception systems still have to represent error, combine evidence and act when confidence is incomplete. It is best read selectively after a first course in probability rather than attempted as a casual cover-to-cover introduction.

Bruno Siciliano and Oussama Khatib, eds., Springer Handbook of Robotics, 2nd ed. (Springer, 2016). Use this as a field atlas rather than a continuous read. Its specialist chapters range from foundations and manipulation to mobile systems, human-robot interaction, medical robotics, field robots and applications. The size can be intimidating, but it is useful for seeing which disciplines sit behind a single deployed machine and for following one interest into authoritative technical literature. Choose one application chapter, then trace its references back to the foundation chapters it depends on.

SRI International, Shakey the Robot historical archive. Use the archive as original evidence of the moment perception, planning and physical action were joined in one mobile research system. Shakey worked in a simplified world and moved slowly, which is part of its value. The material shows how many current autonomy problems were visible by the late 1960s, and how much progress has come from better sensing, computation and integration rather than from changing the basic questions. Pair it with modern rover or warehouse footage and the continuity becomes hard to miss. The archive is short enough for a newly interested reader and concrete enough to puncture vague stories about sudden machine intelligence.

Notes and Sources

The Whole Thing in One Page and Why You Should Care

Definition and categories. The working definition follows ISO 8373:2021 and the International Federation of Robotics: a robot is a programmed actuated mechanism with a degree of autonomy for locomotion, manipulation or positioning. The text uses the boundary as a practical guide rather than pretending that every edge case is clean. Fully teleoperated devices, fixed automation and externally guided vehicles can use robotic technology without satisfying every statistical definition used by IFR.

Industrial scale. IFR's World Robotics 2025 reported 542,000 industrial robots installed during 2024 and an operational stock of 4.664 million, 9 per cent higher than the previous year. Asia accounted for 74 per cent of new installations. These figures refer to industrial robots under standardised definitions and should not be extended to every automatic machine.

Amazon fleet. Amazon announced deployment of its millionth robot in June 2025. The number is a company-reported operational milestone and is used to demonstrate fleet scale, not as an independently audited census of warehouse robotics.

Perseverance. NASA and JPL descriptions of AutoNav state that the rover builds local terrain models, identifies hazards and plans around obstacles while human teams provide broader goals and routes. NASA reported that AutoNav crossed the Snowdrift Peak boulder field, more than 1,700 feet wide, in about a third of the time comparable earlier-rover operations would have taken. In February 2026 JPL reported the first uses of Mars Global Localization, which matched panoramic navigation-camera imagery with orbital maps and located the rover within about 25 centimetres in roughly two minutes.

The Core Ideas

Task and operating domain. The task-first treatment follows the established robotics distinction between nominal capability and performance inside a defined environment and application. ISO safety standards and robotics textbooks both treat the robot as part of a larger system whose tooling, layout, workpiece and human procedures determine behaviour.

Morphology and mechanics. Degrees of freedom, serial and mobile mechanisms, end effectors, compliance, actuators and transmission trade-offs follow Lynch and Park; Craig; Siciliano et al.; and the Springer Handbook of Robotics. The discussion of humanoids makes no forecast of dominance or failure. It treats human form as one response to human-built infrastructure with substantial mechanical costs.

Perception and estimation. The distinction between measurement and state, together with calibration, coordinate frames, fusion, odometry, localisation and SLAM, follows Thrun, Burgard and Fox and standard robotics practice. The text does not imply that all systems use one probabilistic architecture.

Actuation and contact. Electric, hydraulic and pneumatic actuation, transmission effects, torque-speed and thermal limits, force control and impedance control follow the standard control and mechanics texts in the bibliography. Contact examples are explanatory and do not substitute for application-specific design calculations.

Kinematics and planning. Forward and inverse kinematics, Jacobians, singularities, configuration space, trajectory generation and planning methods follow Lynch and Park; Craig; Siciliano et al.; and the Springer Handbook. Sampling and optimisation are presented as broad families, not a claim that one method governs all deployed robots.

Control and autonomy. Feedback, PID, feedforward, nested loops and recovery are standard control-system concepts. Autonomy follows the ISO-derived idea of performing intended tasks from current state and sensing without human intervention. The manuscript rejects a universal autonomy ladder because delegation varies by function and operating domain.

Safety. ISO 10218-1:2025 addresses safety requirements for the industrial robot as partly completed machinery. ISO 10218-2:2025 addresses integration, commissioning, operation, maintenance and decommissioning of industrial robot applications and cells. The separation supports the manuscript's claim that a robot arm cannot make a complete application safe by itself.

How It Actually Works

Unimate. The Computer History Museum records Unimate beginning work at General Motors in 1961, following commands stored on a magnetic drum and handling hot die-cast metal. Historical accounts differ in some plant naming and description, so the narrative retains only the stable facts needed for the mechanism.

Stanford Arm and Shakey. Stanford records the Stanford Arm as developed in 1969 and used in 1974 to assemble a Model T water pump. SRI records Shakey as its research subject from 1966 to 1972 and describes it as the first mobile robot able to perceive and reason about its surroundings. Such priority language is retained only with attribution to the institution.

ROS 2. Official ROS documentation describes modular nodes communicating through topics, services and actions. The manuscript calls ROS 2 middleware and an ecosystem, not an autonomy engine, operating system in the ordinary desktop sense or required architecture for every robot.

Service statistics. IFR states that its 2025 service-robot report is based on a sample of 294 suppliers, is not projected to the whole industry and should not be compared directly with earlier reports because sample composition changes. The narrative therefore avoids a global service-robot stock or trend claim.

Human interaction and recovery. The treatment draws on the Springer Handbook of Robotics and human-centred robotics literature. Commercial intervention rates are commonly proprietary, which is why the manuscript uses intervention burden as a lens without supplying a universal benchmark.

What People Get Wrong and Use It

Accuracy and repeatability. These are standard robot-performance and metrology distinctions. Repeatability can be high when a stable systematic offset makes accuracy poor. Complete application error also depends on tooling, calibration, payload, environment and process.

Collaboration. The correction follows the application-level approach in ISO 10218-2:2025. Protective methods such as monitored stopping, speed and separation monitoring, hand guiding and power or force limiting require design and validation for the integrated task.

AI boundary. Machine learning appears only where it changes perception, planning, instruction or control. Detailed architectures, foundation models, cognition and the wider social debate belong to AI in a Hurry. General algorithm design and complexity belong to Algorithms in a Hurry.

Bibliography

Standards and institutional sources

Amazon. “Amazon Launches a New AI Foundation Model to Power Its Robotic Fleet and Deploys Its 1 Millionth Robot.” 30 June 2025.

Computer History Museum. “1961: UNIMATE.” Timeline of Computer History.

International Federation of Robotics. World Robotics 2025: Industrial Robots. Frankfurt: IFR, 2025.

International Federation of Robotics. “Robot Definitions at ISO.” Accessed 11 August 2026.

International Federation of Robotics. “World Robotics 2025: Service Robots.” Frankfurt: IFR, 2025.

International Organization for Standardization. ISO 8373:2021 Robotics: Vocabulary. Geneva: ISO, 2021.

International Organization for Standardization. ISO 10218-1:2025 Robotics: Safety Requirements, Part 1: Industrial Robots. Geneva: ISO, 2025.

International Organization for Standardization. ISO 10218-2:2025 Robotics: Safety Requirements, Part 2: Industrial Robot Applications and Robot Cells. Geneva: ISO, 2025.

NASA Jet Propulsion Laboratory. “NASA's Self-Driving Perseverance Mars Rover Takes the Wheel.” 1 July 2021.

NASA Jet Propulsion Laboratory. “Autonomous Systems Help NASA's Perseverance Do More Science on Mars.” 21 September 2023.

NASA Jet Propulsion Laboratory. “NASA's Perseverance Now Autonomously Pinpoints Its Location on Mars.” 18 February 2026.

Open Robotics. ROS 2 Documentation: Concepts and Interfaces. Accessed 11 August 2026.

SRI International. “Shakey the Robot.” Historical archive.

Stanford University. “Stanford's Robotics Legacy.” 16 January 2019.

Modern works

Craig, John J. Introduction to Robotics: Mechanics and Control. 4th ed. Harlow: Pearson, 2017.

Lynch, Kevin M., and Frank C. Park. Modern Robotics: Mechanics, Planning, and Control. Cambridge: Cambridge University Press, 2017.

Siciliano, Bruno, Lorenzo Sciavicco, Luigi Villani, and Giuseppe Oriolo. Robotics: Modelling, Planning and Control. London: Springer, 2009.

Siciliano, Bruno, and Oussama Khatib, eds. Springer Handbook of Robotics. 2nd ed. Cham: Springer, 2016.

Thrun, Sebastian, Wolfram Burgard, and Dieter Fox. Probabilistic Robotics. Cambridge, MA: MIT Press, 2005.