Robot

A robot is a programmable machine that can sense its surroundings, decide what to do, and act on the physical world, often with little or no human intervention. The term is both broad and contested: standards bodies define industrial robots narrowly as reprogrammable manipulators, while popular usage also includes mobile, legged, aerial, and humanoid machines, and software agents are usually called bots instead. The field that designs, builds, and operates robots is robotics, which draws on mechanical engineering, electronics, computer science, and control theory. Robots now assemble cars, move goods in warehouses, assist in surgery, and explore Mars, and their development has been shaped by the Industrial Revolution, cybernetics, and, more recently, artificial intelligence.

This entry first traces the word's origin and the long history of automata, then explains the foundations of the field: the mechanical structure, actuation, sensing, and control that let a machine act in the world. It goes on to examine how learning-based methods have changed robot autonomy and to survey the main robot types and their use in manufacturing, medicine, services, and hazardous environments. The later sections cover humanoid robots and their current state, the economic effects of automation, the safety standards and ethical questions that accompany deployment, and the technical problems that remain open.

Etymology and definitions

The word robot comes from Czech. Karel Čapek, a Czech writer, invented the word in his cautionary play R.U.R., which he wrote in 1920 and which the stage performed in 1921. In that play, the robots are not metal machines but biologically engineered humanoids, mass-produced to be efficient and obedient. The Czech noun behind the term points to compulsory labour, and its Slavic root is tied to the idea of bondage. Authorship of the coinage is disputed. Čapek reportedly first wanted a Latin word for work, and his brother Josef Čapek suggested roboti instead. Some accounts therefore credit Josef, while Karel Čapek is the usual attribution in reference works.

The play spread quickly. It introduced the word "robot" to the English language and to science fiction as a whole. By 1923, it had been translated into thirty languages. Another major source of the vocabulary was Isaac Asimov. He coined the term "robotics" in his 1941 story "Liar!", though he later remarked that he believed then that he was merely using an existing word. The laws that became famous were first stated in his 1942 short story "Runaround". They were introduced in that story, although similar restrictions had been implied in earlier stories. Asimov's stories and the wider science fiction tradition gave the word a shared meaning across many languages, and much of the public picture of robots comes from this fiction rather than from engineering practice.

For technical purposes, the most cited definition is the one in ISO 8373. The International Federation of Robotics bases its use of the term "industrial robot" on this standard's definition, which describes a manipulator that is automatically controlled, reprogrammable, multipurpose, and programmable in three or more axes, and which may be fixed or mounted on a mobile platform for industrial automation. Each qualifier does specific work. Reprogrammable means that the programmed motions can be changed without physical alteration of the mechanical system, while multipurpose means the machine can be adapted to a different application only with physical alteration. The three-axis requirement is practical: two axes are needed to reach any point in a plane, and three are needed to reach any point in space. The International Federation of Robotics uses this definition for its annual statistics on installations and shipments, which means the figures it publishes are built on this particular boundary rather than on a general notion of robotness.

Broader working definitions are looser. Many researchers and textbooks describe a robot as a machine that senses its environment, processes that information, and acts on it, following what is often called a sense-plan-act loop. That framing includes many mobile and social machines that the ISO definition leaves out. Industry statistics also use application-based categories. Under this scheme, industrial robots serve industrial automation, while service robots perform useful tasks for people or equipment outside industrial automation. This split creates awkward cases. Autonomous mobile robots are often used in factories, but they do not meet the ISO definition because they have no manipulation capability and no three axes. If such a platform carries a robotic arm, it becomes a mobile manipulator and would count as an industrial robot.

The boundary cases show why no single definition settles the question. Fixed automation is excluded when it is dedicated to one job: a machine dedicated to a single purpose and not reprogrammable is not a robot in this sense. A reprogrammable device with only two axes also falls outside the standard, since it is not "programmable in three or more axes." Meanwhile, safety standards draw the line differently. ISO 10218 treats the manipulating portions of mobile robots as industrial robots for its own purposes. Software bots sense, process, and act on data without any body, so whether they qualify depends on whether a physical embodiment is required. Autonomous cars are usually classed as vehicles, and prosthetic limbs are worn by people rather than working alongside them. Each of these cases can be argued either way, and the entry therefore treats the term as a family of related meanings rather than one fixed category.

Precursors: automata and mechanical devices before 1900

Long before the word "robot" existed, engineers built machines that moved in programmed ways. Many of the principles were worked out in the Hero of Alexandria workshops of the Greek-speaking Roman period. In his treatises on pneumatics and automata, Hero described theatrical machines driven by falling weights, ropes, pulleys, and hydraulics. Some of them ran sequences of actions from a rotating drum fitted with pegs, which trigger cords in order. This is an early form of fixed, repeatable motion, the same basic idea as a cam.

Medieval engineers extended these methods. In 1206 Al-Jazari, a scholar working in the Artuqid court in Diyarbakır, described water-powered machines in his Book of Knowledge of Ingenious Mechanical Devices. These included a boat carrying a band of automated musicians, whose performance was set by pegs on a rotating cylinder, and a humanoid attendant that served drinks. Floats, valves, and water flow regulated the timing. Al-Jazari's work shows how a mechanism could be reset to produce different sequences, a step toward general-purpose control, though each machine still did one fixed routine.

The eighteenth century produced the most famous mechanical figures. Jacques de Vaucanson's digesting duck, shown in the late 1730s, appeared to eat grain, flap its wings, and excrete. Its movements were reproduced by internal cams and linkages, and the public was persuaded that it digested food, though it did not. The Jaquet-Droz figures, built by Pierre Jaquet-Droz and his son in the 1770s, were more sophisticated. The Writer could be set to produce a short text, and its cams could be replaced to change the message. These machines relied on clockwork gears, springs, and precisely shaped cams, which allowed a single mechanism to imitate complex human actions.

The Industrial Revolution made such mechanisms more practical. Improvements in precision machining and the steam engine, developed by figures such as James Watt, supplied reliable power and repeatable motion for factories. The most important step toward programmability came in textiles. Joseph Marie Jacquard's loom, which he developed in 1804, used chains of punched cards to select which threads were raised for each pass of the shuttle. A card sequence therefore encoded a pattern, and changing the cards changed the output without rebuilding the machine. Charles Babbage later drew on this idea for his analytical engine, and it anticipated the logic of the algorithm and of stored instructions.

These devices were mostly demonstrations, entertainment, or single-purpose production tools. They performed their motions open-loop: they did not sense their surroundings, and they could not detect an obstacle, a misplaced object, or a changed load, and then correct their behaviour. Their programs were fixed in gears, cams, pegs, or cards. The governor that Watt adapted for steam engines did respond to speed, and it prefigured the feedback principles that later became central to control engineering, but it regulated a single variable rather than acting in a varied environment. The gap between these machines and later robots lay in the combination of sensing, decision-making, and adaptive action, which was not achieved until the twentieth century, as discussed in the section on cybernetics and feedback.

Cybernetics and the theory of feedback

The idea that a robot is a regulated system, rather than a fixed mechanism, came largely from cybernetics. In his 1948 book Cybernetics, Norbert Wiener treated control and communication as one problem that applies to organisms and machines alike. The book laid the theoretical foundations for the multidisciplinary field of cybernetics, the study of controlling the flow of information in systems with feedback loops, be they biological, mechanical, cognitive, or social. Wiener's interest grew out of wartime work: he drew on his experiences working with anti-aircraft systems during World War II to develop the concept of feedback loops. Gun-aiming problems were hard to solve with fixed predictions, because the target changed course, so the system had to correct itself continuously as it went.

The engineering roots of this idea are older than Wiener's book. In telecommunications, engineers had used negative feedback to hold amplifier output steady against distortion, and the same principle was applied to servomechanisms. A servo compares the measured output with a desired setpoint and produces a correction proportional to the difference. In its simplest form the controller computes an error $e(t) = r(t) - y(t)$, where $r(t)$ is the reference and $y(t)$ the measured position, and then applies an action $u(t) = K_p\,e(t)$ to reduce that error. This is the mechanism by which a machine can track a moving target or hold a joint at a commanded angle, and it is the core of the feedback control that later robot joints rely on. Wiener's contribution was to show that this same logic describes both mechanical servos and nervous systems, which made feedback a general principle rather than a gunnery trick.

Teleoperation supplied the first practical demonstrations of this logic in a manipulator. At Argonne National Laboratory, Raymond Goertz filed a patent for an early master-slave manipulator in 1949 in order to handle radioactive material. Goertz's team built a purely mechanical device in which the slave tongs move in exact correspondence with a master handle. An operator behind a shielding wall moved the master handle, and the slave jaws followed in synchronism. Goertz also recognized the value of electrically coupling manipulators and laid the foundations of modern tele-robotics and bilateral force-reflecting positional servomechanisms. He first described force reflection in print in 1952, so that the operator could feel the forces acting on the remote arm. Dates for the first design vary between sources: one account places the first Argonne device in 1948, while the laboratory report is dated March 1949.

These machines are not robots in the modern sense. They were machines that translated the movements of a human body into mechanical terms somewhere else, and they had no autonomy: every motion came from a person. Their importance lies in the control architecture. Sensing, comparison, and correction were already present, and the open problem was to put a controller in the loop in place of the human. This is the gap that later sensor-equipped and programmed manipulators would fill, and it explains why the definition of a robot in this entry depends on whether a machine can regulate its own behaviour rather than only transmit a human's motion.

Cybernetic ideas framed robots as systems that regulate themselves through sensing and correction, and this view remains central to control engineering. Wiener formalized the field in 1948, identifying feedback as the fundamental mechanism for self-regulating systems. The idea spread through the interdisciplinary Macy conferences (1946-1953), which applied information and feedback principles across biology, engineering, and the social sciences. Modern practitioners often describe the field more narrowly. Today, the Cybernetics program is often narrowly interpreted as being mainly about feedback control systems, although Wiener's broader aim also included understanding intelligence as a system. That wider aim influenced later work on neural networks and computing, and it links cybernetics to the later history of artificial intelligence in robotics.

The first industrial robots

The first commercially used industrial robot grew out of the work of George Devol, an American inventor who became interested in factory automation while working on magnetic recording equipment. His 1954 application described an arm with six degrees of freedom that could hold a sequence of digital instructions on a drum or similar medium, so that the arm could repeat the same motions without a person guiding each one. The patent was not granted until 1961, and it was titled "Programmed Article Transfer" (filed on December 10, 1954 and granted on June 13, 1961).

The design became a product through a partnership with Joseph Engelberger. A financing arrangement with Consolidated Diesel Electronic (Condec) allowed development to continue under a new division, Unimation, with Engelberger as president. Early prototypes were controlled by vacuum tubes used as digital switches, and later versions moved to transistors. Off-the-shelf parts available in the late 1950s, such as digital encoders, were not good enough for the machine, so the company designed and machined almost every component itself.

The commercial breakthrough came with the Unimate. Devol sold the first unit in 1960, and it was shipped to General Motors in 1961. It was installed at a GM plant in Ewing Township, New Jersey, where it handled red-hot door handles and other such car parts by dropping them into cooling liquid. A gripper on a steel arm meant that workers no longer had to touch freshly cast parts, which made the task safer and more consistent. Engelberger's company then sold Unimate-type arms to other manufacturers. Industry-linked histories report that a large number of units were used in die-casting within a short period, although these figures come from promotional sources and are not independently verified.

Sources disagree about when the first industrial robot appeared. Some date it to the 1954 design, which established the basic concept of a programmable multi-jointed manipulator. Others point to the 1961 installation, which was the first production use on a factory line. A few accounts go further back: one industry history places a prototype on a GM die-casting line in 1959, earlier than the 1961 date most histories use. The dispute reflects a choice about what counts as "first": the invention, the patent, the prototype, or the first working deployment. Dates for the company also vary. Unimation is dated to 1956 in some accounts and to 1962 in others, depending on whether the partnership or the formal company is meant.

Adoption spread slowly outside the United States. In the 1970s and 1980s, Japanese and European manufacturers introduced industrial robots into their own plants, mainly for spot welding and painting, where the tasks were repetitive and the paint or weld paths could be fixed in advance.

Most of these early machines were essentially blind and pre-programmed. They had no cameras or force sensors to tell them where a part actually was, so they relied on fixtures and carefully positioned parts to make each motion work. Any change in the layout, part shape, or position usually required reprogramming or new tooling. This limited their use to fixed, structured tasks, and it explains why sensing and adaptive control became major research goals in the decades that followed. The same constraint is also why the earliest robots were most valuable in large factories where production volumes justified custom installations.

Early research in artificial intelligence and mobile robots

Shakey, built at the Artificial Intelligence Center of Stanford Research Institute (now SRI International) from 1966 to 1972, was the first mobile robot to reason about its actions. The project was partly motivated by military interest, since the Advanced Research Projects Agency, the predecessor of DARPA, was interested in finding ways to use robots and artificial intelligence for military reconnaissance. Shakey's cameras let it perceive its environment, interact with humans via natural language, plan its activities, and take actions to achieve assigned goals. Its navigation relied on search: it planned waypoints while avoiding obstacles, and the search method developed for this work, known as A*, became a widely used algorithm. The robot itself was never a practical machine, but it advanced visual analysis, route finding, and object manipulation. Funding ended in the early 1970s; Nilsson later recalled that the project was renewed through 1971, after which the sponsor stopped supporting robot work.

Stanford's vehicle work followed a separate path. The Stanford Cart was originally built by James L. Adams to study the feasibility of controlling a lunar explorer from Earth, and graduate students extended it over two decades. Rod Schmidt's 1971 doctoral thesis addressed vision and control for a robot operating outdoors. The cart could follow an unbroken white line on a road for about 15 meters before breaking track. Hans Moravec later obtained stereo images by moving a black and white video camera side to side to create a stereo baseline. The planning method depended on a model of the surroundings built from camera images, and the plan changed as the Cart perceived new obstacles on its journey. Speed was very low: the Cart moved 1 m every 10 to 15 min, in lurches. Outdoors, the same algorithms could carry it only about 15 meters before shadows created false edges to track.

Moravec's results showed that a three-dimensional model of the environment was valuable for path planning, and they also exposed how hard perception becomes outdoors. Stereo remained computationally prohibitive at the time, while rangefinders were too slow for the task. His broader observation, that abilities humans find effortless, such as perception and sensorimotor coordination, are the hardest for machines to reproduce, is now often called Moravec's paradox. His 1980 Stanford thesis dealt with obstacle avoidance and navigation in the real world by a seeing robot rover, and this line of work influenced the mobile robots that followed.

Progress slowed as expectations outran results. Minsky and Papert's 1969 book on perceptrons was widely read as showing neural networks were fundamentally limited. In 1973, Sir James Lighthill's report to the UK Science Research Council concluded that AI had failed to deliver on its promises. It judged that robot research on eye-hand coordination was "entirely disappointing". The report contracted funding for AI research in British universities sharply, and in the United States DARPA's cutbacks to academic AI research in general followed in 1973–74. Hardware limits compounded the problem, since insufficient computing power, difficulty tackling complex problems, results below expectations, and growing skepticism from funders converged during the 1970s. The term "AI winter" itself first appeared in 1984 as the topic of a public debate at the annual meeting of AAAI. Further cuts followed in the late 1980s, including the 1988 cancellation of new spending on AI by the Strategic Computing Initiative. The practical effect was that general-purpose autonomy, meaning robots that could act reliably in unstructured and changing surroundings, advanced far more slowly than the early demonstrations had suggested. This constraint shaped the later move toward machine learning and probabilistic methods.

Mechanical structure and kinematics

Most manipulators are open chains of rigid links joined by revolute (rotary) or prismatic (sliding) joints. The first link is fixed to a base, and the last link carries a tool or gripper, called the end effector. Each joint adds motion of its own, and the number of these independent motions, the degrees of freedom (DOF), sets how far the arm can reach and how many ways it can orient its tool. A rotary joint contributes exactly one degree of freedom, and its axis of rotation defines one axis of the coordinate frame for that link. A six-DOF arm can in principle place its tool at any position and orientation within its workspace, whereas a three-DOF arm can position a tool but cannot control its orientation independently. Extra joints add flexibility, including the ability to route around obstacles, but they also add mass, cost, and control complexity.

Forward and inverse kinematics

Forward kinematics computes where the end effector is, given the joint angles. The method multiplies a transformation matrix for each link, so the result is a single pose (position and orientation) with respect to the base. This direction is straightforward and deterministic: one set of joint angles produces one pose.

Inverse kinematics runs the other way. It starts from a desired tool pose and solves for the joint values that achieve it. This problem is much harder. The governing equations are nonlinear, and a given target may have several solutions, one solution, or none if it lies outside the reachable workspace. For six-DOF arms, algebraic elimination can produce extremely large polynomial systems, so controllers usually rely on closed-form solutions for arms with particular geometries, such as those whose wrist axes meet at a point, or on iterative numerical methods. The latter typically use the Jacobian matrix, which relates joint velocities to tool velocities. At singular configurations the Jacobian loses rank, and the arm temporarily cannot move in some direction no matter how the joints are driven. Controllers must detect and avoid these postures.

Conventions and parallel mechanisms

To describe a serial chain consistently, engineers commonly use the Denavit–Hartenberg (DH) convention, introduced in 1955. It assigns a coordinate frame to each link and characterises every link–joint pair with four parameters: two describing link geometry and two describing joint motion. Because each link reduces to a standard set of parameters, the forward-kinematics chain can be computed systematically and the same notation can be used across many robot designs.

Parallel mechanisms take a different structural approach. In a delta robot, for example, several light arms connect a fixed base to a small moving platform, and the actuators sit on the base rather than on the moving parts. Keeping the moving mass low allows rapid acceleration, which is why delta arms are common in high-speed pick-and-place work. The trade-off is a smaller workspace and a more complicated relationship between actuator motion and platform motion.

Structural trade-offs

Structural design balances stiffness, mass, workspace, and precision, and these factors cannot all be optimised at once. Stiffer links resist deflection under load and improve accuracy, but they usually weigh more. Heavier links raise the moment-of-inertia of the arm and increase the gravitational torque that the joints near the base must supply, which in turn requires larger motors and more energy. Long reach expands the workspace but magnifies deflection at the tool. Lighter, more compliant structures are safer around people and cheaper to move, yet they can be less precise and may vibrate more. These choices therefore shape not only what a robot can do but also how safely it can operate and how much power it consumes.

Actuation and power

Actuators convert stored energy into motion, and the type chosen for each joint determines a robot's strength, speed, precision, and how it behaves on contact. Most designs use one of four families: electric, hydraulic, pneumatic, or series-elastic variants of these. Power supply, usually a battery for mobile machines, then sets how long the robot can work.

Electric motors

Electric motors are the most common actuators, mainly because a motor's torque can be regulated with relatively simple electronics, which makes precise joint positioning practical. Motors spin quickly but produce modest torque for their size, so they are usually paired with gearboxes or harmonic drives that trade speed for torque. A harmonic drive achieves a large reduction in a compact package with almost no play in the gears, though it is costly and can be sensitive to shock. Most industrial arms that weld, paint, or assemble parts rely on motor-driven joints for this reason.

Hydraulic and pneumatic actuators

Hydraulic actuators push pistons with pressurized oil. Hydraulics have impressive power to weight compared to electric motors. This makes them suitable for heavy payloads and for legged machines that must push hard against the ground. The drawbacks are the pump, valves, hoses, and fluid leaks that add weight and maintenance, and they are often more difficult to use in applications requiring good force control, such as legged robots, due to their poor backdrivability. Boston Dynamics' earlier Atlas illustrated this trade-off by using hydraulic joints in its legs, torso, and upper arms, with electrically driven forearms and wrists.

Pneumatic actuators use compressed air. They are inexpensive, light, and naturally yielding, so they are often used in robot hands and artificial muscles, where modest forces are acceptable. Because air is compressible, pressure changes lag behind valve commands, which makes precise position and force control harder than with motors.

Series elastic and variable stiffness actuators

A series elastic actuator (SEA) places a spring between the motor and the output, so the controller can estimate force from the spring's deflection. Using Hooke's law, the output torque is approximately $\tau = k\,\Delta x$, where $k$ is the spring stiffness and $\Delta x$ is the deflection. Engineers have explored SEAs since the 1990s as a way to make force control more faithful. Such an elastic element gives SEAs tolerance to impact loads, low mechanical output impedance, passive mechanical energy storage, and increased peak power output. These properties improve safety when a robot touches people, and they help legged machines absorb the shock of landing. The main cost is bandwidth: actuators are built to be as stiff as possible to increase the bandwidth. Designers therefore must size the spring to carry the load without making it so rigid that the compliance is lost.

Variable stiffness actuators go further by changing the stiffness of the joint during operation, for example by altering a spring preload or the lever arm acting on it. This allows a single joint to be stiff for accurate positioning and soft for safe contact. Hydraulic versions also exist: these actuators are similar in operation to the electrical Series Elastic Actuators except that a hydraulic cylinder replaces the motor and ball screw.

Energy storage and battery limits

Mobile robots draw their energy from batteries, most often lithium-ion cells. The key figure is energy density, the amount of energy stored per unit mass or volume. A heavier battery adds load that the actuators must then supply, so energy density and actuator efficiency interact. Runtime is roughly the usable stored energy divided by the average power drawn by actuators, sensors, and computing hardware. Because motors and hydraulic pumps convert energy with losses, the same stored energy can yield very different working times depending on the actuator choice. For this reason, power consumption often determines how long a robot can operate between charges, and it is a central constraint in legged and humanoid designs.

Sensing and perception

Robots use sensors in two broad groups. Proprioceptive sensors measure the machine's own state, such as joint angles and body motion. Exteroceptive sensors measure the surroundings, such as obstacles, objects and contact with the environment. Control depends mainly on the first group, while interaction with the world depends on the second.

Internal sensing: joints and body motion

The most common internal sensor is the encoder. An optical encoder mounted on a motor shaft or wheel passes light through a rotating disc with slits, and a photodiode turns the resulting pulses into angle and speed. Its resolution, usually given in cycles per revolution, sets the smallest angle it can detect, and it must respond fast enough for the shaft's rotation rate. Incremental encoders count pulses from a starting point and lose their position when power is cut, which is why resolvers, which give an absolute angle without battery backup, are used in demanding settings.

Inertial measurement units (IMUs) combine accelerometers and gyroscopes to estimate tilt, rotation and acceleration. They are central to balance in legged robots, drones and humanoids. Joint encoders and IMUs typically update at 100 to 1000 Hz, far faster than many camera-based estimators, and they do not depend on lighting, which makes them useful in fog or dust. Joint torque or current sensing adds a further internal signal: motor current can serve as a rough proxy for how hard an actuator is working, which is the basis of many force-control schemes that rely on electric motors and torque feedback.

Contact sensing: force, torque and tactile skin

A six-axis force-torque sensor at the wrist reports the total contact force and torque on the hand, but on its own it cannot say where on the surface the contact occurred. Joint motor measurements can estimate external forces and detect collisions, yet they also cannot localise the contact point. Six-axis sensors are also bulky, costly and rigid, and they are easily damaged by overload, which limits their use in fingertips. One line of research has shown that high-resolution joint torque sensing in a redundant arm can give a robot an intrinsic sense of touch across its whole body, without a separate artificial skin.

Tactile sensors fill the gap at the point of contact. Designs include piezoresistive and capacitive cells arranged as taxels, modular skin patches, and optical visuo-tactile fingers, in which a camera watches the deformation of a soft elastomer. These sensors let a robot detect slip, adjust grip force and handle fragile or deformable objects. Survey work on robot learning reports that conditioning control policies on force feedback enables skills such as pouring precise volumes, inserting pegs into tight holes and grasping delicate items, which are difficult or impossible without touch.

Exteroceptive sensing: cameras, depth sensors and lidar

Cameras provide dense, rich appearance data at low cost, but a monocular camera is very sensitive to illumination, and monocular depth estimation remains imprecise: some state-of-the-art RGB-based methods show errors above 50 cm on indoor benchmarks. Stereo and structured-light depth cameras give direct range data over short distances, and lidar, which measures distance with laser pulses, gives precise geometry over longer ranges. In one comparison of sensors for fruit harvesting, a stereo depth camera covered roughly 0.6 to 6 m, while a lidar covered about 0.2 to 20 m with error under 2 cm, though with lower resolution than the camera. The choice of sensor is therefore a trade-off among range, accuracy, resolution, cost and robustness to lighting, and the computer vision pipeline built on top of these data must handle each sensor's failure modes.

Sensor fusion and calibration

Robots combine these streams through sensor fusion. Fusion can occur at the level of symbolic estimates, at the level of extracted features, or at the level of raw data, and each choice trades accuracy against computational cost. Practical fusion requires spatial calibration, which finds the rigid transformation between sensor frames, and temporal alignment, usually done with timestamps so that IMU, camera and depth samples correspond to the same moment. Fusing lidar and camera data is difficult because their point densities, ranges and viewpoints differ, so calibration and registration algorithms are an active research topic. Fused estimates are then used for localisation, obstacle detection and, in human–robot collaboration, safety monitoring.

Limits in unstructured settings

Perception in unstructured settings, such as recognising objects under changing light, clutter, occlusion or reflective and transparent surfaces, remains one of the main bottlenecks for autonomy. Vision can locate contact but cannot measure force and is often blocked by occlusion, while force sensors measure force but not position. Simulating tactile sensors is also hard, particularly the shear forces produced by soft sensor skins. Industrial robots that work in highly structured cells can rely on precise fixtures and predictable geometry, but robots that must adapt to changing environments need perception that combines several modalities and still functions when individual sensors degrade.

Control, planning, and localization

Most robot behaviour is produced by a layered stack. Low-level controllers track joint references, planners choose the motion, and estimators work out where the robot is. Each layer depends on the others, and an error in one tends to show up in the rest.

At the lowest level, control loops compare a measured joint angle with a desired one and command the electric-motor to reduce the difference. Proportional–integral–derivative (PID) controllers are the most common choice. The command is built from three terms:

Here $e(t)$ is the tracking error. The proportional term responds to the present error, the integral term removes steady offsets such as those caused by gravity, and the derivative term damps oscillation by reacting to how fast the error is changing. PID controllers remain popular because they need little knowledge of the robot's dynamics and can be tuned by hand or automatically. Their weaknesses are that the derivative term amplifies sensor noise and that gains tuned for one pose may perform poorly in another, since the effective moment-of-inertia of an arm changes as it moves. These limits motivate model-based methods higher in the stack.

Motion planning asks how the robot should move from a start state to a goal state without hitting anything. Planners usually work in configuration space, where each point is a complete set of joint values (or base position and heading for a mobile robot). Obstacles become forbidden regions in this space, and the remaining region is the free space. Sampling-based planners approximate this free space with random samples. A probabilistic roadmap (PRM) samples collision-free configurations, connects nearby ones, and then searches the resulting graph with methods related to shortest-path search and graph-theory; it is suited to repeated queries in the same environment. Rapidly-exploring random trees (RRT) instead grow a tree outward from the start configuration, one random sample at a time, and check each new connection for collisions; RRT is a single-query method and is usually faster for one-off problems. Variants such as RRT* rewire the tree to approach cost-optimal paths. Optimization-based planners instead write path length, smoothness, clearance from obstacles, and joint or torque limits as a cost with constraints, then solve for a trajectory with mathematical-optimization techniques. Sampling methods cope well with high-dimensional and cluttered spaces but do not optimize cost by default, while exact planning becomes intractable as the number of degrees-of-freedom grows.

Simultaneous localization and mapping (SLAM) addresses the case in which a mobile robot has neither a map nor a reliable global position, as is common indoors, where GPS is unavailable. The robot must estimate its pose while building the map, and each estimate depends on the other: a drifting pose corrupts the map, and a distorted map misleads the pose estimate. Early formulations used probabilistic filters such as the extended Kalman filter or particle filters, in which the map and pose are estimated jointly and uncertainty is propagated from step to step. Many modern systems instead represent the problem as a graph whose nodes are robot poses and landmarks and whose edges are sensor measurements, and then solve a large nonlinear least-squares problem to make the whole trajectory consistent. Practical SLAM pipelines combine a front end, which extracts features from lidar or camera data and associates them across frames, with a back end that performs the optimization. Loop closure is especially important: when the robot recognizes a previously visited place, the accumulated drift can be corrected across the whole trajectory. The method is standard in autonomous vehicles, warehouse robots, and drones, and it is also used in augmented reality.

Model predictive control (MPC) takes a different approach from fixed feedback gains. At each control step, it uses a dynamic model of the robot to solve a finite-horizon optimal control problem that minimizes tracking error and effort while respecting constraints such as torque limits, joint limits, and friction at contact points. It then applies only the first part of the resulting control sequence and repeats the calculation with the new measured state. This receding-horizon scheme handles constraints explicitly, but it is computationally demanding, and early implementations were restricted to slow motions. Faster solvers and parallel hardware have made real-time use practical. Whole-body control extends this idea to the entire robot. Rather than controlling joints independently, it computes torques for all joints together so that several prioritized tasks, such as keeping the centre of mass over the support feet, placing a hand at a target, and maintaining posture, are satisfied as far as the contact constraints allow. For legged robots, this is harder because the base is not fixed to the ground and contact forces change as feet touch and lift. Whole-body MPC for a humanoid was demonstrated on hardware as early as 2015. More recently, research groups have shown real-time whole-body MPC on quadrupeds and full-size humanoids using the iterative linear-quadratic regulator with the MuJoCo physics engine, and sampling-based predictive control has been run on a legged robot in real time.

Learning-based autonomy and embodied AI

Reinforcement learning treats control as a decision problem. A policy maps the robot's observations to actions, and the learner adjusts that mapping to maximise the expected sum of rewards, commonly written as $J(\theta)=\mathbb{E}_{\pi_\theta}\left[\sum_{t}\gamma^{t} r_t\right]$. The robot acquires its skill through trial and error, usually within a Markov decision process framework. Because physical trials are slow and can damage hardware, most training runs in simulation, where thousands of virtual robots can practise in parallel. A well-known example is OpenAI's 2019 work on a dexterous robot hand that learned to solve a Rubik's Cube after training entirely in simulation with randomised physical parameters. Designing the reward function remains difficult, since a poorly specified reward can lead to behaviour that scores well but does not accomplish the intended task.

Imitation learning offers a different route. Instead of specifying a reward, the designer records demonstrations, often collected through teleoperation or by physically guiding the arm, and trains a model to reproduce the demonstrated actions. The simplest version, behaviour cloning, is a supervised regression problem solved with gradient-descent. It is data-efficient for narrow tasks, but small errors can accumulate when the robot drifts into states that the demonstrations never covered. Practical systems therefore combine demonstrations with reinforcement learning fine-tuning, or collect corrective data during operation. The approach reduces the need for hand-written control programs, but it shifts the cost to data collection, which must include a diverse range of objects, lighting conditions, and failure cases.

Vision-language-action models extend these ideas to general-purpose behaviour. A single network receives camera images and a natural-language instruction and outputs motor commands, often by representing actions as tokens in the same way that a large language model represents words. Google DeepMind's RT-2, presented in 2023, is a widely cited example: it was built on a vision-language model pre-trained on web-scale data and then trained on robot trajectories, which allowed it to respond to some instructions involving objects and concepts absent from its robot training set. Several other large AI laboratories and robotics companies have since pursued similar architectures. These models draw on computer vision, self-supervised pre-training, and transfer learning, and they are sometimes described as foundation models for embodied agents.

Two obstacles limit how far learned policies generalise. The first is the sim-to-real gap: simulators approximate contact, friction, sensor noise, and actuator delays imperfectly, so a policy that performs well in simulation may fail on hardware. Techniques such as domain randomisation, careful system identification, and fine-tuning on real data reduce this gap but do not eliminate it. The second is data scarcity. Text and images on the internet are abundant, whereas robot data must be gathered physically, one trial at a time, and is expensive to produce. Shared datasets that pool recordings from many laboratories and robot types have been assembled to address this, yet they still cover a small fraction of the situations a general-purpose robot would encounter. As a result, learned policies often succeed in the conditions represented in their training data and degrade in unfamiliar environments, a limitation that continues to shape research on robot autonomy.

Types of robots

Robots are usually grouped by physical form and by the setting they were built for. The categories overlap, because one machine can combine several of them, such as a wheeled base carrying a manipulator arm.

The most numerous group is the fixed industrial arm. Mounted on a floor or in a workcell, it performs welding, painting, machine tending, and material handling, and it accounts for most robots installed in factories. Traditionally these arms have run inside fences or behind light curtains, which keep people out of the zone where a fast, forceful arm moves.

Collaborative robots, or cobots, were designed to share a workspace with people. The four collaborative operating methods named in ISO/TS 15066 are a safety-rated monitored stop, hand guiding, speed and separation monitoring, and power and force limiting (PFL). Power and force limiting is the mode most people mean when they say "collaborative robot." In this mode the robot detects contact, commonly through torque sensing or current monitoring, and keeps the force it applies below limits set for each body region. For the torso, the figure most often quoted for quasi-static contact is 150 newtons. The cost of this approach is speed: because contact force is capped, a fast arm cannot be used freely near people, so throughput-heavy jobs often switch to speed-and-separation monitoring or full guarding. Cobots suit small-batch manufacturing, where frequent reprogramming and changeover matter more than cycle time.

Mobile robots include automated guided vehicles (AGVs) that follow floor markers or magnetic guides, warehouse robots that carry shelves or containers to pickers, and autonomous delivery vehicles. Earlier systems depended on fixed markers or prebuilt maps. Newer ones rely more on onboard sensing and computer-vision, and they increasingly use mapping and localization methods that let them work in changing surroundings. Autonomous cars sit at the edge of this category, and whether they count as robots depends on the definition in use (see automobile and electric-vehicle).

The remaining categories are more varied:

• Legged and humanoid robots must balance continuously while moving, which requires controlling many degrees-of-freedom at once.
• Aerial drones are limited mainly by how long their battery can supply power for flight.
• Underwater vehicles are either remotely operated or autonomous, and are used for pipeline inspection and seabed mapping.
• Soft robots are built from compliant materials, which makes them safer around people but harder to model and control precisely.
• Exoskeletons are worn on the body to assist movement, with applications in rehabilitation and in lifting support for workers.
• Microrobots operate at very small scales and are explored for targeted medical treatment and laboratory research, an area connected to medicine.

These categories describe typical designs rather than strict boundaries. A mobile manipulator, for example, is both a mobile robot and an arm, and a cobot can be mounted on an AGV. Classifying a machine often depends on the task it performs more than on its shape.

Manufacturing and logistics applications

The automotive industry was the first large-scale user of industrial robots, and it remains the biggest single customer. Its work is spot and arc welding, painting, body assembly, screw fastening, and end-of-line inspection. Industry statistics show how concentrated this demand still is. According to the International Federation of Robotics (IFR), about 40% of new industrial robot installations in the United States in 2024 went to car makers, and electric-vehicle production lines contributed to a 10.7% rise in US automotive installations that year. Globally, 542,000 industrial robots were installed in 2024, and China alone accounted for roughly 54% of them. Automotive demand is cyclical, however. In the European Union, automotive installations fell 25% in 2025, to 14,900 units, although the sector still made up about a quarter of the market. The automobile industry's dominance is therefore strong but not steady.

Electronics, metal processing, and food packaging have grown in importance because their tasks are repetitive and high-volume. In electronics, robots place components, load and unload test fixtures, and handle small parts where consistent positioning matters more than human dexterity. Metal plants use robots for machine tending, press loading, deburring, and grinding, where the robot works continuously beside a cutting or forming machine. Food packaging uses robots for picking, placing, cartoning, and palletizing. IFR data for the United States show that adoption in the food sector rose about 30% in 2025, reaching roughly the same number of installations as metal and electronics. Food and other soft or irregular products remain harder to handle than rigid parts, so robots in this sector rely on specialised grippers and vision systems.

Warehousing and e-commerce fulfilment represent the fastest-changing part of the field. Mobile robots carry shelves or totes to human pickers or to stations, which reduces walking time inside large facilities. Automated storage and retrieval systems use shuttles and grid-based racks to move goods without manual travel. Robotic picking arms are used to take individual items from bins and place them into orders. Picking is the difficult step: products vary in shape, size, and packaging, so the arm needs computer vision and grasp planning to succeed reliably. Most fulfilment centres therefore combine several kinds of automation, with mobile robots handling transport and fixed arms handling selected tasks at the edges of the workflow.

The cost of a robot is rarely the cost of the whole system. Integration includes cell design, fixtures, programming, connection to conveyors and factory software, safety fencing and risk assessment, and commissioning. Integrators put the arm at roughly 25–50% of a typical project, although estimates vary. One widely cited planning rule puts total installed cost at about two to five times the hardware price, and integration labour alone is often given as 150 to 400 hours per cell. Figures of this kind differ between vendors and integrators, so they should be read as planning ranges rather than fixed values. Safety validation and tooling are often underestimated, and fixture work can overrun budgets when part presentation proves more variable than expected.

These cost structures explain where robots are adopted. Automation pays off most when a product is made in large runs with stable geometry, because the integration effort can be spread across many units. Manufacturers that change products often, or that run small batches, face repeated reconfiguration costs, which is one reason collaborative robots, standardised cells, and leasing or robot-as-a-service models have spread among smaller firms. In warehouses, fleet software and integration with inventory systems can cost more than the vehicles. Adoption therefore depends less on the robot's technical capability than on whether its total installed cost can be recovered through labour savings, higher throughput, or fewer errors.

Medical, rehabilitation, and surgical robots

Medical robots are one of the few robot categories in which a human professional usually stays in control of the task, and clinical settings impose unusually strict limits on design, testing, and approval. The best-known example is the da Vinci system made by Intuitive Surgical. Its surgeon console holds a stereoscopic viewer, two master tool manipulators that the surgeon grips, and foot pedals that switch control modes. A separate patient-side cart carries the arms that hold the endoscope and the instruments. The instruments have wristed joints, so their tips gain at least six degrees of freedom, and the surgeon's hand movements are translated into those motions. The machine does not choose the cut; it carries out the surgeon's intent through a small incision, which is why the approach is called minimally invasive.

Regulation shapes how these systems reach hospitals. The first FDA-cleared surgical robot was AESOP, a camera-positioning device cleared in 1993 as a moderate-risk Class II device through the 510(k) pathway. A 510(k) submission must show that a new device is substantially equivalent to a legally marketed predicate device, which makes incremental changes easier to clear than fundamentally new automation, which may need the more demanding premarket approval route. A 2024 systematic review of 49 surgical robots cleared by the FDA between 2015 and 2023 classified most of them as Level 1 systems that only assist the surgeon, with a small share reaching conditional autonomy. Competition is growing: Medtronic's Hugo system received FDA clearance for urologic procedures, and it had already held a CE mark in Europe since 2021.

Wearable powered exoskeletons form the second group. These devices have motorized hip and knee joints and were first designed to let people with complete paralysis from spinal cord injury stand and walk; they are now also used in gait training for stroke survivors. Evidence is strongest for spinal cord injury. A 2015 systematic review of 15 studies examined the walking speeds people reached with powered exoskeletons and how those speeds related to factors such as the level of injury and the number of training sessions. A more recent study of 74 participants with spinal cord injury, each completing at least five training sessions with a commercial device, reported reduced spasticity, with larger gains after 15 or more sessions. For stroke, the evidence is more mixed. One small University of Texas study found that gait training did not significantly improve walking performance in five participants, although wearing the device changed their muscle activity patterns, and a separate comparison reported that the exoskeleton group regained a more controlled neuromuscular rhythm than the control group.

Laboratory automation is the third major area. Robotic systems handle sample tubes, pipetting, and specimen preparation in research laboratories and diagnostic facilities. They increase throughput and reduce the variation and errors that come with repetitive manual handling, which matters when large volumes of samples must be processed consistently.

Across all these uses, adoption depends on more than technical performance. Medical robots must demonstrate safety and effectiveness, and regulators often accept evidence from comparisons with established devices and real-world clinical data, as the FDA did when it cleared the Senhance system in 2017. Hospitals must also justify the cost through clinical outcomes and reimbursement, train surgeons and staff, and maintain sterile, reliable operation over long service lives. These requirements make medical robotics slower to deploy than industrial automation, even where the underlying engineering is mature.

Service, domestic, and social robots

Domestic robots

Consumer robots are among the most widely used robots in homes, and robotic vacuum cleaners are the clearest case. Early models bounced around rooms in largely random paths. Current models use sensing to build maps: one market analysis reports that over 72% of robot vacuum models launched since 2020 integrate multiple navigation technologies such as infrared and LiDAR. Coherent market research describes how LiDAR-based navigation and SLAM algorithm integration let these devices clean in systematic sequences rather than random patterns. Many models also return to a docking station when their battery runs low, so they can finish scheduled cleaning without intervention.

Battery capacity limits what these machines can do. One report notes that most robotic vacuum cleaners operate between 60–120 minutes, which may be too short for larger homes. Market figures for the sector also vary widely between analysts. Adoption claims are often supplied by manufacturers, such as the widely quoted figure that about 70% of American homes own one, which comes from a single market player and should be treated as indicative rather than definitive.

Companion and social robots

Companion robots are designed for emotional engagement rather than chores. Sony's AIBO was the best-known early example. Sony announced a prototype Aibo in mid-1998, and the first consumer model was introduced on 11 May 1999. Its second generation added sensors and LEDs, along with speech recognition and image capture. The dog could walk, sit, and respond to voice commands or touch, but it had difficulty recognizing and responding to a wide range of vocal commands due to processing limitations. Sony ended production in January 2006, citing disappointing sales, and one account puts total sales at about 150,000 units over the product's commercial life. Its high price also limited its reach.

In therapeutic settings, the pet-like robot Paro became a reference point. A 2003 study by Wada and colleagues reported that Paro could lower stress and build a bond with older adults, even though it had limited emotion expression and no emotion perception or speech abilities. Later social robots are more complex. Humanoid platforms such as Pepper and NAO combine speech, gesture, and emotion recognition. Pepper is a 1.20 m tall teenager-like humanoid robot, and NAO is 57 cm tall, with an appearance designed to be more acceptable to users. Research prototypes go further: one system recognizes emotional states from facial expression and speech sentiment and adjusts its responses accordingly. Speech interaction remains fragile in practice. Researchers note that the prevalence of local dialects among elderly populations makes standard speech recognition models less reliable.

Service robots

Service robots in care facilities are often justified by their practical value. Service robots for senior care are gaining popularity because they can improve the quality of life for the old while decreasing the workload on caregivers. Some monitoring tools avoid cameras altogether. Wi-Fi sensing devices can detect movement through changes in wireless signals, and infrared systems can flag unusual activity patterns for caregivers, which reduces the need for visual surveillance.

Privacy and user expectations

Social and domestic robots raise questions about what users expect and what data they collect. A 2025 study of socially assistive robots in elder care found that stakeholders distinguished several kinds of privacy concern. Informational privacy refers to the right of individuals to control the disclosure of their personal information to third parties, which matters most when a robot talks about residents. Physical privacy is a separate issue, and residents differed on whether a robot should simply knock before entering a room or ask for permission. The same study reported that stakeholders prioritized privacy over individuality due to the perceived risks of data breaches and a lack of trust in the robot's capabilities. These concerns connect to wider debates about ethics and data privacy in machines that observe people in their homes.

Robots in extreme and exploratory environments

Robots are often sent where people cannot safely work or survive. The same constraints recur in space, the deep ocean, and damaged nuclear plants: communication links are weak, sensing is limited, and human operators must supervise from a distance.

Planetary rovers

Mars received its first wheeled rover in mid-1997. Sojourner, which weighed about 10.6 kilograms, reached the surface by driving down a ramp that the lander's solar panels formed. It consisted of a lander, renamed the Carl Sagan Memorial Station, and a lightweight wheeled robotic Mars rover. The landing itself relied on inflated airbags rather than retro-rockets. The probe bounced along the surface until it came to rest. Sojourner had been planned for about seven sols of work, yet it operated for 83 sols and covered roughly 100 metres of terrain. Its semi-autonomous "behaviors" made it a favorite of the public.

The mission ended because the rover depended on the lander for its link to Earth. Contact was lost on 27 September 1997 when the Pathfinder lander's battery failed. Later vehicles grew out of this design lineage. NASA describes Sojourner as the forerunner of more advanced rovers, including Curiosity and Perseverance.

Underwater vehicles

Two kinds of underwater robot matter most here. Tethered remotely operated vehicles (ROVs) are cabled to a ship and remain the standard tool for handling objects. Inhabited submersibles and ROVs remain the only option for manipulation tasks such as sampling, deploying and recovering experiments on the seafloor. Autonomous underwater vehicles (AUVs) are untethered and follow a pre-set route. Underwater, a GPS fix is unavailable, so an AUV estimates its position from arrays of acoustic beacons on the seafloor, or from a combination of Ultra Short Base Line acoustic communication, GPS positioning, and inertial navigation when it is near the surface. Below the surface it depends on dead reckoning from depth, inertial and Doppler velocity sensors.

AUVs have been in use since the 1960s, and by the 1980s, AUVs were also being developed to specialize in deep-sea mapping. Their limits are practical. They have a maximum range of about 1,000 meters before their high-frequency sound gets absorbed, and they move at slow speeds to conserve power. Some designs combine both modes: hybrid vehicles (HROVs) can do both, roaming freely for survey and tethering for sampling.

Nuclear and disaster sites

Robots entered the nuclear field after the 1979 accident at Three Mile Island. Carnegie Mellon researchers built robots in the late 1970s and early 1980s to inspect and perform repairs in the basement of the plant. The 2011 Fukushima Daiichi disaster was far larger. After the tsunami the loss of power meant that the cooling systems failed, leading to the fuel melting. Shortly afterwards, two PackBot ground robots from iRobot entered Unit 1 and Unit 3 and performed readings of temperature, oxygen levels, and radioactivity. The data revealed high levels of radioactivity, too high for humans to access the facilities.

These machines had clear limits. Radiation can disrupt radio signals, so the operators used fibre-optic tethers. The Packbots and Warriors could be tele-operated from up to 220 meters and 500 meters, respectively. Neither model was designed for extreme heat or submersion. Cleanup work has continued for more than a decade. TEPCO is deploying a remote-controlled robot to extract melted fuel debris from Unit 2, and an estimated 2,850 highly contaminated sandbags are being removed from the plant's basements, with a target of completing sandbag extraction by March 2028.

Disaster response and search and rescue

Search-and-rescue robots face similar problems in a different setting. After the attacks on the World Trade Center, robots inspected areas beneath debris and entered confined spaces that humans or dog rescuers could not enter. Operators worked under stressful conditions, and minimum sensor information, caused by dust or poor lighting, led to mistakes in robot operation. Rubble and unstructured terrain therefore push research on mobility, sensing, and teleoperation, since machines must act reliably while the human supervising them has incomplete information.

Humanoid robots: development and the 2026 state

Humanoid robots have a longer research history than their recent publicity suggests. Honda's P-series prototypes came first. The P2, introduced by Honda in 1996, was a self-contained autonomous bipedal humanoid capable of stable dynamic walking and stair-climbing. Honda's ASIMO followed, with development beginning in 1999 and the robot unveiled in October 2000. Its walking relied on Zero Moment Point control, a method that keeps the ground-reaction point inside the support area so the robot does not tip. Later models could also climb stairs and recognize voices and faces. Honda retired the line in 2022. ASIMO's main lesson was that stable walking was largely a control problem, and later humanoid programmes have shifted toward hands, dexterous manipulation and whole-body coordination.

Boston Dynamics' Atlas shows the move from demonstration to product. At the Consumer Electronics Show in Las Vegas, the company unveiled the product version of its new Atlas robot, and the fully electric humanoid was revealed during Hyundai's global CES media day presentation. The company said it would begin production at its Boston headquarters immediately, and that all Atlas deployments are already fully committed for 2026, with fleets scheduled to ship to Hyundai's Robotics Metaplant Application Center and Google DeepMind. Outside customers are expected later, since the company plans to add additional customers in early 2027. Its planned training approach leans on new AI foundation models for a wide variety of industrial tasks, beginning in the automotive sector, which connects to foundation models and reinforcement learning. Hyundai has also tied the programme to a factory design capacity of 30,000 robots a year, as part of a $26 billion investment in the United States.

Market figures for 2026 are still preliminary and disagree. IDC estimates that global humanoid robot shipments approached 25,000 units in the first half of 2026, up 432.1 percent year on year, and reports revenue exceeding $740 million. Other trackers counted fewer units: Smart Analytics Global put the total at roughly 19,100 units, while Counterpoint estimated 22,000. These gaps most likely reflect different counting and estimation methods. Shipments also do not show how many units are doing productive work after delivery. Production is concentrated in China: IDC attributes about 77.9% of the global market share to Chinese shipments, and Agibot shipped more than 8,600 humanoid robots in the first half of 2026. IDC's longer-term outlook is more aggressive, projecting global shipments exceeding 750,000 units by 2030.

Deployment is spreading beyond showcases. IDC says the market is increasingly expanding into areas including industrial manufacturing, retail, logistics, tourism and consumer applications. Agibot reported that it reached 20,000 cumulative robots produced in September 2026, and that in September, Agibot and Chimelong began large-scale deployment of more than 300 robots at Chimelong Spaceship Park in Hengqin. Manufacturing itself is changing: according to one contract manufacturer, humanoid robots are now entering a phase where production volumes are measured in the tens of thousands, and hand assembly is no longer viable.

Whether humanoids compete with other automation depends on reliability, energy use and cost. Each joint adds actuators, sensors and wiring, and the mobile platform must carry its own power, so battery capacity and energy density set practical limits on runtime. Atlas's commercial path is still early, and Boston Dynamics has a long road ahead of it. The wider questions about how such machines will affect labour markets and productivity remain open, and the embodied AI approaches that many developers now pursue have yet to be proven across many sites and tasks.

Economic effects and labor markets

Robots raise productivity in manufacturing and logistics, where they can increase output, improve quality, and lower the cost of some goods. Their broader economic effect depends on how quickly firms adopt them and how the gains are shared, so economics treats robotization as a question of technological-change and productivity rather than a simple replacement of workers by machines.

The best-known empirical work concerns US local labor markets. Daron Acemoglu and Pascual Restrepo measured robot exposure across US commuting zones for 1990 to 2007, using adoption in other advanced economies as an instrument. They found large and robust negative effects of robots on employment and wages, estimating that one more robot per thousand workers reduces the employment-to-population ratio by between 0.18 and 0.34 percentage points, with a wage decline of between 0.25 and 0.5 percent. Their model separates two channels. Greater robot penetration lowers wages and employment through a displacement effect, as robots take over tasks workers previously performed, but it can also raise them through a productivity effect, as other industries expand. Later reanalyses have disputed the linear form of this result. One study reports a U-shaped relationship between robotization and employment in the most robust regressions, so that beyond a given level of robotization, further increases induce an employment rise. Other work links robotization to wage polarization, finding that robotisation has significantly contributed to wage polarisation in the US between 1990 and 2007. The evidence therefore varies by industry, region, occupation, and period, and claims about the net effect on employment or the shift toward higher-skilled work remain contested.

Installation data show that robot deployment is highly concentrated. The International Federation of Robotics (IFR) reports that the global operational stock of industrial robots increased 9% to a record 5 million units in 2025, and the 11% jump in annual installations meant factories worldwide installed more than 600,000 new units. China dominates. Its installations reached 354,000, up 20%, accounting for 59% of global deployments. The rankings below China have shifted: the United States, with nearly 38,500 installations, moved ahead of Japan into second place globally. Japan recorded 36,219 installations, down 19%, falling to third place, while South Korea installed 30,000 units, down 1%, remaining the fourth-largest market. Germany had fewer than 25,000 installations, down 8%, ranking fifth globally and first in Europe. The IFR expects annual installations to rise 9% to 655,000 units in 2026, which would continue the growth in installations, although the mix of countries driving it is changing.

Investment in robotics has grown quickly, most visibly in humanoid machines. Estimates differ considerably because trackers use different definitions and cut-off dates. Dealroom data cited by one industry brief report humanoid robotics startups had raised $8.7 billion in venture funding through July, already nearly double the total for all of 2025. Another Dealroom-based summary gives $8.6 billion so far this year, already 1.8 times more than all of 2025, and a more recent Dealroom snippet puts year-to-date venture investment at $11 billion. Broader robotics totals are higher still: one tracker reports that robotics startups had raised roughly $18.8 billion globally by June 22, already exceeding the total for all of 2025. Observers warn that capital is running ahead of operations. One industry analysis argues that the industry still has to prove the machines can work reliably and cheaply enough to compete with workers and simpler automation. Whether this investment translates into large-scale labor displacement or into new kinds of work is therefore still an open question.

Safety standards, ethics, and governance

Safety rules for robots are organised by type of machine rather than through one general code. For industrial robots, the core reference is ISO 10218. The revised ISO 10218 contains two parts: 10218-1:2025, aimed at manufacturers, and 10218-2:2025, aimed at system integrators. These documents replace the 2011 versions of Parts 1 and 2. In North America, the earlier edition had been adopted as ANSI R15.06 in the United States and CSA Z434 in Canada.

The 2025 revision changed several things. It made functional safety requirements more explicit, so that they are stated rather than implied. It introduced new robot classifications with corresponding functional safety requirements and test methodologies. It also brought in cybersecurity. The updated standard requires measures against unauthorized access, manipulation and data loss. The collaborative-robot rules were absorbed as well: integrated safety requirements for collaborative robot applications now consolidate the previously separate ISO/TS 15066.

Collaborative robots, which share space with workers, follow one of four operating modes. These are safety-rated monitored stop, speed and separation monitoring, hand guiding, and power and force limiting. Speed and separation monitoring relies on keeping a computed distance between the machine and a person, so the robot slows or halts as that distance shrinks. It prevents collisions by imposing a speed-dependent separation distance between the human and the robot. Power and force limiting takes a different approach by accepting that contact may occur. It ensures that the robot's kinetic energy stays below limits that could injure a person on contact. Among the four modes, power and force limiting is the only one that allows sustained physical contact between a moving robot and a human operator. Vendor guidance summarising ISO/TS 15066 gives, for example, a maximum quasi-static force of 150 N on the torso as a biomechanical limit. Combining the two strategies is still a research topic rather than settled practice.

Service and personal care robots fall under a different standard. ISO 13482 specifies requirements and guidelines for the inherently safe design, protective measures, and information for use of personal care robots, in particular mobile servant robots, physical assistant robots, and person carrier robots. Its scope has deliberate limits: it does not apply to robots travelling faster than 20 km/h, robot toys, water-borne or flying robots, industrial robots covered by ISO 10218, robots that are medical devices, or military and public force application robots. The 2014 edition has remained in force; the publication was last reviewed and confirmed in 2020, and therefore this version remains current. The ISO 8373 vocabulary standard supplies the shared terms that regulators and manufacturers use across these documents.

Standards address physical hazards far more readily than ethical ones. One analysis notes that the ISO 13482 standard sets out a number of safety, but not ethical, hazards, including battery charging, robot motion, incorrect autonomous decisions, and lack of awareness of robots by humans. A new generation of explicitly ethical standards is now emerging. Asimov's Three Laws of Robotics, introduced in his fiction, are not used as engineering requirements, but they remain a common reference point in public debate about how robots should behave around people.

Several governance questions remain open. Privacy is a central concern for camera-equipped and home robots that collect data about the people around them. Liability is unclear when an autonomous system causes harm, because responsibility may be spread across manufacturers, integrators, software developers and operators. Worker consultation matters where robots change tasks and schedules on the factory floor. These questions connect to existing frameworks on data privacy, cybersecurity, law, ethics and human rights, and they are being addressed through policy and standard-setting in many jurisdictions, but no single international regime yet covers all robot types.

Open problems and technical limits

Dexterous manipulation is among the least solved problems in the field. Rigid grippers and multi-fingered hands built from electric motors and linkages can grasp boxes and bottles reliably, but they struggle with cables, cloth, thin paper, and other objects whose shape changes on contact. Compared with rigid object manipulation, handling deformable objects is considerably more complex and remains an open research problem. Solving it requires progress in hardware, sensing, deformation modelling, planning, and control at the same time. Simulators add a further difficulty: modelling contacts for soft and deformable objects requires a level of realism that existing simulators cannot yet deliver. Robot hand design is therefore an active research area, with compliant and underactuated fingers, tactile skins, and learning-based control all being tested as partial answers.

Power is a second constraint, and it bites hardest on mobile and humanoid machines. A legged robot pays a continuous energy cost to stay balanced, even when it is idle. Onboard computing adds to the load: running large foundation models on GPUs draws significant power, rapidly depletes batteries, and generates concentrated heat near electronic components. The energy available is limited by pack mass as well as by energy density in lithium-ion cells. One industry guide notes that a humanoid's torso can accommodate a pack of only about 5-6 kg. Reported runtimes vary widely. One vendor-oriented guide puts typical operation between 2 and 4 hours, another cites 90–120 minutes per charge for some models, and manufacturer figures differ again. These numbers are not directly comparable, because the robots, tasks, and test conditions differ, so buyers need duty-cycle data from the intended workflow. Thermal limits compound the problem: GPUs require active cooling, and sustained motor loads generate heat that degrades efficiency and component lifetime. Battery cycle life also falls under heavy use, since high-frequency start-stop and high-rate discharge conditions can accelerate battery degradation.

Generalization is the third limit. Learned control policies, often trained with reinforcement learning in simulation, transfer imperfectly to hardware, and the sim-to-real gap is one reason. Another is that generalizing acquired knowledge to actions that differ from those previously demonstrated remains difficult. Reliability in the field is therefore harder than reliability in the lab. Commentators describe the target for commercial deployment as very high uptime; one source puts customer expectations at industrial-grade reliability (95-99% uptime), which many demonstration systems do not reach. Human supervision is often hidden in the results. One buyer's guide reports that research sessions with a small humanoid are often limited to 30-60 minutes due to thermal management, safety supervision requirements, and the cognitive demand on teleoperation operators.

The economic question depends on these technical limits. Whether robots can replace or augment human labour depends on reliability over long periods, maintenance costs, and the expense of integrating a system into a particular site. These costs are often underestimated in early demonstrations, and the gap between a convincing demonstration and a dependable production system is where many projects stall. This is why adoption is usually measured in productivity gains and labour-market effects only after systems have run unsupervised for long periods, rather than in showcase tasks.

References

1. R.U.R.britannica.com
2. Rossum’s Universal Robots by Karel Čapek (1920)psychogeographicreview.com
3. 101 years ago: origins of the word 'robot'therobotreport.com
4. R.U.R.en.wikipedia.org
5. Isaac Asimov and the Three Laws of Roboticsscihi.org
6. Three Laws of Roboticsen.wikipedia.org
7. Industrial Robotsifr.org
8. 1 Introduction 25 1.2 Industrial robots - definition and classificationifr.org
9. Industrial roboten.wikipedia.org
10. What is an industrial robot? Industrial robot definitionrobotnik.eu
11. What are the Key Factors Used to Classify Industrial Robots?digikey.com
12. ISO/DIS 8373 - Robotics -- Vocabularystandards.iteh.ai
13. Who did invent the word "robot" and what does it mean? - Adelaide Robotics and Computer Science Academy (ARCSA)roboticsacademy.com.au
14. Mutation Testing for Industrial Robotic Systemsarxiv.org
15. Cybernetics or Control and Communication in the Animal and the Machinemitpress.mit.edu
16. Norbert Wiener The Founder Of Cyberneticsquantumzeitgeist.com
17. Raymond Goertzen.wikipedia.org
18. MASTER-SLAVE MANIPULATORscispace.com
19. How the Atomic Age Gave Us Robot Surgeonsgizmodo.com
20. When control meets large language models: From words to dynamicsarxiv.org
21. Principles and Practice of Deep Representation Learning: or a Mathematical Theory of Memoryarxiv.org
22. Master-Slave Manipulator (Technical Report)osti.gov
23. From sci-fi to reality: a trip through robotics evolutioninnovacion.uas.edu.mx
24. 1954 – ElectroMechanical Manipulator – Ray Goertz (American) - cyberneticzoo.comcyberneticzoo.com
25. Norbert Wiener and the Birth of Cybernetics: The Science of Feedbackmedium.com
26. NIHF Inductee George Devol Invented the Industrial Robotinvent.org
27. Unimateen.wikipedia.org
28. George Devolen.wikipedia.org
29. Unimationen.wikipedia.org
30. Joseph Engelberger and Unimate: Pioneering the Robotics Revolutionautomate.org
31. Unimate Industrial robotshvachko.net
32. Shakey - CHM Revolution - Computer History Museumcomputerhistory.org
33. Fifty years of Shakey, the "world's first electronic person"newatlas.com
34. SRI’s AI Center celebrates 60 years of AI impact - SRIsri.com
35. Stanford’s Robotic Historystanfordmag.org
36. 1960 - Stanford Cart - (American) - cyberneticzoo.comcyberneticzoo.com
37. Read "Technology Development for Army Unmanned Ground Vehicles" at NAP.edunationalacademies.org
38. [PDF] The Stanford Cart and the CMU Roversemanticscholar.org
39. Hans Moravecen.wikipedia.org
40. History of the AI Winters: AI History Guideaitimeline.world
41. Deep Dive: The 1973 Lighthill Report: How One Mathematician Accidentally Triggered AI's Dark Agejrdelaney.substack.com
42. Lighthill Report on Artificial Intelligence (1973)achievements.ai
43. AI winteren.wikipedia.org
44. Artificial Intelligence: Seventy Years of History, Bringing It to Business - INNOVATING NEWSen.innovando.news
45. Shakey the Robot - SRI Internationalsri.com
46. SRI's Pioneering Mobile Robot Shakey Honored as IEEE Milestone - IEEE Spectrumspectrum.ieee.org
47. Rover Visual Obstacle Avoidancemlanthology.org
48. The AI Winter: Why AI Almost Died (Twice) - Beginners in AIbeginnersinai.org
49. Forward Kinematicsopentextbooks.clemson.edu
50. Robot Kinematicsfiveable.me
51. Inverse Kinematics for a 6-Degree-of-Freedom Robot Manipulator Using Comprehensive Gr\"obner Systemsarxiv.org
52. (PDF) Forward and Inverse Kinematic Analysis of Robotic Manipulatorsacademia.edu
53. Modeling and Application of Series Elastic Actuators for Force Control Multi Legged Robotsarxiv.org
54. Series Elastic Actuator: Design, Analysis and Comparisonintechopen.com
55. Design of robust controller applied for series elastic actuators in controlling humanoid's jointarxiv.org
56. Development of Bidirectional Series Elastic Actuator with Torsion Coil Spring and Implementation to the Legged Robotarxiv.org
57. (PDF) Series elastic actuators for high fidelity force controlresearchgate.net
58. Proprioceptive Sensorsieee-dataport.org
59. Robotics Module 6# Perception in Action - Robotics Sensorsdilipkumar.medium.com
60. Principles of Robot Autonomyarxiv.org
61. Proprioceptive Sensors: The Robot's Inner Eye - DataRoot Labsdatarootlabs.com
62. Proprioceptive Invariant Robot State Estimationarxiv.org
63. Curved patch mapping and tracking for irregular terrain modeling: Application to bipedal robot foot placementarxiv.org
64. ShanghaiTech Mapping Robot is All You Need: Robot System for Collecting Universal Ground Vehicle Datasetsarxiv.org
65. Probabilistic Multimodal Depth Estimation Based on Camera-LiDAR Sensor Fusionarxiv.org
66. [2307.10519] Probabilistic Multimodal Depth Estimation Based on Camera-LiDAR Sensor Fusionarxiv.org
67. Accurate Fruit Localisation for Robotic Harvesting using High Resolution LiDAR-Camera Fusionarxiv.org
68. SLCF-Net: Sequential LiDAR-Camera Fusion for Semantic Scene Completion using a 3D Recurrent U-Netarxiv.org
69. A LiDAR-depth camera information fusion method for human robot collaboration environment - ScienceDirectsciencedirect.com
70. Singular Value Decomposition (SVD) Method for LiDAR and Camera Sensor Fusion and Pattern Matching Algorithmncbi.nlm.nih.gov
71. Tactile Robotics: An Outlookarxiv.org
72. Intrinsic sense of touch for intuitive physical human-robot interactionscience.org
73. Customizing Textile and Tactile Skins for Interactive Industrial Robotsarxiv.org
74. Towards Forceful Robotic Foundation Models: a Literature Surveyarxiv.org
75. GelSlim 4.0: Focusing on Touch and Reproducibilityarxiv.org
76. Tactile Robotics: An Outlookarxiv.org
77. SLAM - Simultaneous localization and mappingsbg-systems.com
78. An Introduction to Simultaneous Localization and Mapping (SLAM) for Robots - Technical Articlescontrol.com
79. Simultaneous localization and mappingen.wikipedia.org
80. Active SLAM: A Review On Last Decadearxiv.org
81. Mapping and tracking system with features in three-dimensional spaceimage-ppubs.uspto.gov
82. Whole-Body Model-Predictive Control of Legged Robots with MuJoCoarxiv.org
83. Real-Time Whole-Body Control of Legged Robots with Model-Predictive Path Integral Controlarxiv.org
84. (PDF) Whole-Body Model Predictive Control for Biped Locomotion on a Torque-Controlled Humanoid Robotresearchgate.net
85. Motion planning in high-dimensional spacesarxiv.org
86. Sampling-Based Motion Planning: A Comparative Reviewarxiv.org
87. HAS-RRT: RRT-based Motion Planning using Topological Guidancearxiv.org
88. Collaborative Robot Safety Standards: ISO 10218 & R15.06inmotion.global
89. Power and Force Limiting Cobots (October 2026 The Complete Guide)smashingrobotics.com
90. Collaborative robot safety standards you must know - Standard Botsstandardbots.com
91. Standardizing Collaborative Robots: What is ISO/TS 15066? - Engineering.comengineering.com
92. Collaborative Robot Safety: ISO/TS 15066 Guide (2026)evsint.com
93. Cobot Safety: ISO/TS 15066 and Power-and-Force-Limiting in Practicereemanbot.com
94. U.S. Companies Invest Heavily in Robots - IFR Preliminary Results - International Federation of Roboticsifr.org
95. Robot Installed in US Auto Industry Up by Double Digits - International Federation of Roboticsifr.org
96. International Federation of Robotics (IFR) Secretariat President Jane Heffnerifr.org
97. US Robot Industry Returns to Double Digit Growth - International Federation of Roboticsifr.org
98. Automotive leads new installs: IFR data for Europe and world 2024 - News - portal Plastech.bizplastech.biz
99. Industrial Robot Installations 2024: 542,000 IFR Robotics Statsmanufacturingleadgeneration.com
100. The True Cost of an Industrial Robot: the 2-5x Ruleindustrialroboticshub.com
101. Industrial Robot Cost: 2026 Price Guide by Payloadthebotscout.com
102. Full article: The role of lower limb exoskeletons in rehabilitation: a scoping reviewtandfonline.com
103. Wearable powered overground exoskeleton reduces dose-dependently the spasticity in individuals with spinal cord injurynature.com
104. Gait speed using powered robotic exoskeletons after spinal cord injury: a systematic review and correlational studyncbi.nlm.nih.gov
105. How can powered exoskeletons restore walking ability after stroke?physicsworld.com
106. Frontiersfrontiersin.org
107. Patient-side surgeon interface for a minimally invasive, teleoperated surgical instrumentimage-ppubs.uspto.gov
108. Imaging cannula with a hinged tipimage-ppubs.uspto.gov
109. Surgical port featureimage-ppubs.uspto.gov
110. Levels of autonomy in FDA-cleared surgical robots: a systematic reviewnature.com
111. Robot Autonomy for Surgeryarxiv.org
112. FDA clears new robotically-assisted surgical device for adult patientsfda.gov
113. Medtronic’s Hugo surgical robot earns FDA clearancemedtechdive.com
114. Robot Vacuum Cleaner Market Size, Sharebusinessresearchinsights.com
115. Residential Robotic Vacuum Cleaner Market Trends, 2026-2033coherentmarketinsights.com
116. Robotic Vacuum Cleaners Market Size, Share & Growth360researchreports.com
117. AIBOen.wikipedia.org
118. Early Robot Dog from 2000s: A Robotic Pet Historytherobotacademy.com
119. Whatever Happened To Sony's Aibo Robot Dog? - SlashGearslashgear.com
120. Artificial Emotional Intelligence in Socially Assistive Robots for Older Adults: A Pilot Study - PMCpmc.ncbi.nlm.nih.gov
121. Employing Socially Assistive Robots in Elderly Caredl.acm.org
122. Social robot in service of the cognitive therapy of elderly people: Exploring robot acceptance in a real-world scenario - ScienceDirectsciencedirect.com
123. Robots for Elderly Care: Review, Multi-Criteria Optimization Model and Qualitative Case Study - PMCpmc.ncbi.nlm.nih.gov
124. Which Values Matter to Socially Assistive Robots in Elder Care Settings? Empirically Investigating Values That Should Be Embedded in SARs from a Multi-Stakeholder Perspectivearxiv.org
125. Aibo, Sony's robot dog could come backrobot-advance.com
126. Robotic Vacuum Cleaner Market Size, Share, Trends and Forecast by Type, Type of Charging, Distribution Channel, Application, End User, and Region, 2026-2034researchandmarkets.com
127. Residential Robotic Vacuum Cleaner Market Size & Trends, 2033persistencemarketresearch.com
128. Robotic Vacuum Cleaners Market Size, Share and Forecast 2032credenceresearch.com
129. Personalized Cognitive Support via Social Robotsmdpi.com
130. Image and Speech Recognition Technology in the Development of an Elderly Care Robot: Practical Issues Review and Improvement Strategiesncbi.nlm.nih.gov
131. Mars Pathfindernasa.fandom.com
132. Mars Pathfinder Lander Prototypeairandspace.si.edu
133. Sojourner Rover — The First Mars Roverorbitalradar.com
134. 25 Years Ago: Mars Pathfinder Launches to Mars to Deploy Sojourner, the First Planetary Rover - NASAnasa.gov
135. Techniques for Deep Sea Near Bottom Survey Using an ...robots.stanford.edu
136. Autonomous Underwater Vehicles (AUVs): Their past, present and future contributions to the advancement of marine geoscience - ScienceDirectsciencedirect.com
137. Autonomous Underwater Vehicles - NOAA Ocean Explorationoceanexplorer.noaa.gov
138. Surface-based sonar system could rapidly map the ocean floor at high resolutionnews.mit.edu
139. Underwater vehicleswhoi.edu
140. Robots Arrive at Fukushima Nuclear Site with Unclear Missionscientificamerican.com
141. Detection of Simulated Fukushima Daichii Fuel Debris Using a Remotely Operated Vehicle at the Naraha Test Facilityncbi.nlm.nih.gov
142. robots enter fukushima reactors detect high radiationspectrum.ieee.org
143. Meltdown Maneuvers: Advanced Robotics for Safer Nuclear Cleanup - Tech Insightseepower.com
144. Robots to remove 2,850 radioactive sandbags from Fukushima plantinterestingengineering.com
145. Mixed-Initiative variable autonomy for remotely operated mobile robotsarxiv.org
146. Sojourner (rover)en.wikipedia.org
147. Sojourner — The first successful Mars roverspace.com
148. Emergency response to the nuclear accident at the Fukushima Daiichi Nuclear Power Plants using mobile rescue robotsresearchgate.net
149. 30 Years Ago, Robots Learned to Walk Without Falling - IEEE Spectrumspectrum.ieee.org
150. ASIMOen.wikipedia.org
151. Asimo - ROBOTS: Your Guide to the World of Roboticsrobotsguide.com
152. Boston Dynamics Unveils New Atlas Robot to Revolutionize Industrybostondynamics.com
153. Atlas Humanoid Robots Production ‘Fully Committed’ For 2026, Factory Will Build 30,000 Per Yearforbes.com
154. Agibot leads global humanoid robot shipments in H1 2026, IDC data showsroboticsandautomationnews.com
155. Global humanoid robot shipments surge 432% in H1; China accounts for nearly 80%, Agibot overtakes Unitree for top spot — BigGo Financefinance.biggo.com
156. Humanoid Robot Shipments Jump 432%, China Has 95% Market Sharefinance.yahoo.com
157. Institutions: China's humanoid robot shipments exceeded 19,000 units in H1, accounting for about 77.9% of the global totalpanews.io
158. IDC: In the first half of the year, China shi...chaincatcher.com
159. Humanoid Robots Transition From Hand Assembly to Mass Production - QUE.comque.com
160. Milestone-Proposal:Honda's P2, First Bipedal Robot, 1996 - IEEE Milestones Wikiieeemilestones.ethw.org
161. Robots and Jobs in the US Labor Marketnber.org
162. NBER WORKING PAPER SERIES ROBOTS AND JOBS: EVIDENCE FROM US LABOR MARKETSnber.org
163. Robots are not always bad for employment and wages - ScienceDirectsciencedirect.com
164. Robots and wage polarisationcepr.org
165. 5 million robots are now at work in factories worldwide, reports the IFR - The Robot Reporttherobotreport.com
166. IFR Reports Global Industrial Robot Fleet Reaches 5M as China Drives Growththeaiinsider.tech
167. The State of Humanoid Robotics in 2026: Trends, Challenges and Opportunitiestheaiinsider.tech
168. Top 10 humanoid robot startups to watch in 2026, ranked by total funding — TFNtechfundingnews.com
169. Humanoid Robotics Investment in 2026: Record $16.2B Funding, Top Companies, ETFs, and Ant Group's Robbyantstudioglobal.ai
170. Industrial Revolution: 500,000+ Robots Installed in 2024dcvelocity.com
171. Humanoid robotics, in chartsdealroom.co
172. Updated ISO 10218automate.org
173. News: Updated ISO 10218: Major Advancements in Industrial Robot Safety Standards Now Availableautomate.org
174. Standard Update Alerttuv.com
175. ISO 10218 industrial robot safety standard receives major overhaultherobotreport.com
176. Implementing Speed and Separation Monitoring in Collaborative Robot Workcellsresearchgate.net
177. ISO 13482 - Robots and robotic devices - Safety requirements for personal care robotsstandards.globalspec.com
178. ISO 13482:2014 - Robots and robotic devices — Safety requirements for personal care robotsiso.org
179. Robot Accident Investigation: a case study in Responsible Roboticsarxiv.org
180. Challenges and Outlook in Robotic Manipulation of Deformable Objectsarxiv.org
181. Trends and challenges in robot manipulationscience.org
182. Humanoid Factors: Design Principles for AI Humanoids in Human Worldsarxiv.org
183. Humanoid Robot Battery Selection: Battery Life, Capacity, and Future Trendscmbatteries.com
184. Whose Leading The Way In Robotics, Humaniods & Physical AI in 2026raisesummit.com
185. Complete Guide to Battery Selection for Humanoid Robots in 2026lybatt.com
186. Humanoid Robotics Challenges [2026] - Blog - RoboZapsblog.robozaps.com
187. Humanoid Robot Buyer's Guide 2026: Compare Models, Costs & Use Casesroboticscenter.ai
188. PP-Tac: Paper Picking Using Tactile Feedback in Dexterous Robotic Handsarxiv.org
189. Learning thin deformable object manipulation with a multi-sensory integrated soft handarxiv.org
190. Future of Humanoid Robots: What 2026 Is Provingblog.robozaps.com