A robot is a machine that senses the world around it, decides what to do, and then moves to do it. Those three steps are what separate a robot from an ordinary machine. A toaster heats bread the same way every time no matter what. A robot vacuum notices a chair leg, decides to steer around it, and drives a different path than it did yesterday.
Why robots surprise people
The robot of movies and cartoons is a metal person with a head, arms, and legs. Almost none of the robots working today look like that.
Most robots in a factory are arms bolted to the floor. They weld car doors, spray paint, and lift heavy parts. More than 4.6 million industrial robots were working in factories around the world in 2024. Not one of them has a face.
Robots built for the home are even more common than that. Close to 20.1 million of them were sold in 2024, and 97 percent were built for household chores.
Here is the other surprise. Things that feel easy to you are the hardest things for robots. A computer beat the world chess champion back in 1997. Nearly thirty years later, research robots can fold laundry and pick up objects they have never seen before, but no machine does either job as reliably as you do. Unfamiliar objects and messy rooms are still one of the biggest unsolved problems in robotics. Walking across a cluttered room, catching a ball, telling a sock from a shirt: your brain does all of that without effort, and engineers still cannot match it.
Key facts about robots
The word robot comes from a 1920 play by the Czech writer Karel Capek. His brother Josef suggested it.
It comes from robota, a Czech word for hard forced labor.
The word existed for 41 years before the first real factory robot.
The first factory robot was called Unimate. It started work at a General Motors plant in 1961, lifting hot metal parts.
More than 4.6 million industrial robots were working worldwide in 2024.
Close to 20.1 million robots for home use were sold in 2024, and 97 percent of them do household chores.
The Roomba, released in 2002, was the first home robot that millions of people actually bought.
A robot helicopter named Ingenuity flew on Mars on April 19, 2021, the first powered flight on another planet.
Ingenuity was supposed to fly five times. It flew 72 times over nearly three years.
Robots help surgeons operate, holding tiny tools with no shake at all.
NASA sent a robot called Robonaut 2 to the International Space Station in 2011.
Robots that go where people cannot
Some of the best robot jobs are the ones no person could do.
Mars is the clearest example. Radio signals take between 4 and 24 minutes to travel from Earth to Mars, one way. That means nobody can drive a Mars rover with a joystick: by the time you saw a rock and pushed the controls, half an hour would have passed and the rover would already have hit it. So the rovers drive themselves. NASA’s Perseverance rover builds a three-dimensional map of the ground ahead while its wheels are still turning, spots hazards, and steers around them on its own.
Robots also go into volcanoes, under the ocean, into damaged buildings, and inside nuclear power plants. The rule is simple: send the machine where sending a person would be dangerous, boring, or impossible.
Common myths about robots
Myth: Robots are always shaped like people.
The two most common kinds are a factory arm with no head or legs and a flat disk that vacuums a floor. Robots shaped like people are rare and mostly still experimental.
Myth: Robots think like people do.
A robot follows instructions and patterns. It does not understand what it is doing the way you understand what you are doing.
Myth: Isaac Asimov’s Three Laws of Robotics are built into real robots.
Those laws come from a story published in 1942. They are fiction. Real robot safety uses fences, speed limits, force limits, and emergency stop buttons.
Myth: A robot that falls over is broken.
Modern walking robots are designed to get back up on their own after a fall.
Frequently asked questions about robots
Where does the word robot come from?
From a play called R.U.R., written in 1920 by Karel Capek. The Czech word robota means forced labor.
What was the first robot?
The first industrial robot was Unimate, which started work in a New Jersey car factory in 1961. It was a heavy arm that moved hot metal parts.
How does a robot know where it is?
Sensors. A robot arm has a device at each joint that reports the exact angle. A robot vacuum uses cameras or lasers to build a picture of the room.
Are there robots in space?
Yes. Rovers on Mars, a helicopter that flew there, robotic arms on the space station, and a robot named Robonaut 2 that was sent up in 2011.
Can robots think for themselves?
Not the way people do. They follow programs and learned patterns. Even the smartest robot today has no idea what it is doing or why.
What job do most robots do?
Housework, if you go by how many machines there are. Close to 20.1 million robots for the home were sold in 2024, and 97 percent of them were built for household chores. In factories, the two biggest buyers of new robots in 2024 were electronics companies and carmakers.
Source notes
The origin of the word is documented in the entry on R.U.R., Capek’s 1920 play, and the first factory robot in the Unimate record. Counts of working robots come from the International Federation of Robotics, which publishes separate yearly summaries for industrial robots and for service robots. NASA’s report on the first Ingenuity flight covers the Mars helicopter, and the home robot story comes from the Roomba entry.
A robot is a machine that senses its surroundings, processes what it senses, and acts on the result. All three parts matter. A dishwasher runs a fixed cycle regardless of what is inside it, so it is not a robot. An arm that finds a part on a conveyor belt, calculates where to grip it, and picks it up is one. More than 4.6 million industrial robots were working in factories worldwide in 2024, and almost none of them look anything like a person.
The word came before the machine
Karel Capek’s play R.U.R., short for Rossum’s Universal Robots, opened in Prague in January 1921 after being written in 1920. Capek needed a name for artificial workers built to do human labor. He first considered labori, from the Latin for work, and thought it too stiff. His brother Josef, a painter, suggested roboti, from the Czech robota, meaning the forced labor a serf owed a landowner.
Two details are usually lost. First, the robots in the play are not machines at all. They are artificial people grown from a chemical paste, closer to what we would now call synthetic biology. Second, the play ends with the robots wiping out humanity, so the word arrived already carrying a warning.
The first real industrial robot did not appear until 1961, 41 years after the word was invented.
From Unimate to Shakey
George Devol filed a patent in 1954 for what he called programmed article transfer: a machine that could be taught a sequence of movements and repeat it. The patent was granted in 1961.
Devol met Joseph Engelberger, an engineer whose stated goal was to build mechanical workers for factories, and the two founded Unimation, the world’s first robotics company. Their machine, Unimate, went to work at a General Motors plant in New Jersey in 1961, pulling castings out of a die-casting machine while they were still red hot. Doing that by hand was dangerous work, which is exactly why it was chosen for a machine.
Unimate repeated a fixed sequence. It could not perceive anything or adjust. The next step came from a research institute in California, where between 1966 and 1972 a team built a wheeled robot named Shakey. Shakey could be given a goal, such as pushing a box off a platform, and work out the steps itself. It got its name from how much it wobbled. Its computing happened on a large computer in another room, connected by radio, because nothing small enough to ride along existed yet.
Key facts about robots
The word robot comes from Karel Capek’s 1920 play, suggested by his brother Josef, from the Czech robota, meaning forced labor.
George Devol patented the first industrial robot design in 1954; it was granted in 1961.
Devol and Joseph Engelberger founded Unimation, the first robotics company.
Unimate began work at a General Motors plant in 1961.
Shakey, built between 1966 and 1972, was the first mobile robot that could plan the steps of a task by itself.
Isaac Asimov stated the Three Laws of Robotics in a 1942 short story called Runaround. He had already coined the word robotics a year earlier, in the May 1941 story Liar!
More than 4.6 million industrial robots were in use worldwide in 2024, with about 542,000 installed that year.
China installs more industrial robots per year than the rest of the world combined.
Honda’s ASIMO, a walking humanoid, was shown in 2000 and retired in 2022.
A cobot, short for collaborative robot, is built with force limits so it can safely share a workbench with a person.
Why walking is harder than chess
In 1988, the roboticist Hans Moravec pointed out something that still holds. Computers find it comparatively easy to do things humans consider hard, such as playing chess or solving equations, and extremely hard to do things a one-year-old does without thinking, such as recognizing a face or walking across a room.
A computer beat the world chess champion in 1997. Almost thirty years later, research robots can fold laundry and pick up objects they have not seen before, but none of them does either job as reliably as a person. Unfamiliar objects and cluttered spaces remain one of the biggest unsolved problems in robotics.
The usual explanation is evolutionary. Vision, balance, and hand control were refined over hundreds of millions of years and run below conscious awareness, so they feel effortless. Abstract reasoning is recent, deliberate, and slow, so it feels hard. The parts that feel hard turn out to be the parts that are easiest to write down as rules.
Common myths about robots
Myth: Asimov’s Three Laws govern real robots.
They are a fictional device from 1942, and most of Asimov’s stories are about the laws producing unintended results. Real robot safety uses physical barriers, speed and force limits, and emergency stops, defined in published international standards.
Myth: Robots are replacing all factory workers.
Robot numbers have roughly doubled in a decade, and manufacturing employment in many countries has not collapsed. What robots take over first are specific tasks that are dangerous, repetitive, or require inhuman consistency.
Myth: A humanoid shape is the goal of robotics.
Most designers avoid it. Legs are harder to balance than wheels and a human shape is rarely the best tool for a given job. Humanoids matter mainly where a robot must work in spaces built for people.
Myth: The first robots were built in the 1980s.
Unimate started work in 1961, and research robots were rolling around laboratories in the 1960s.
Frequently asked questions about robots
Who invented the first industrial robot?
George Devol designed and patented it, and Joseph Engelberger turned it into a business. Their machine, Unimate, started work in 1961.
What made Shakey different from earlier robots?
Earlier robots were told each individual movement. Shakey could be given a goal and break it into steps itself.
Are Asimov’s laws used in real robots?
No. They are fiction. Real safety comes from standards such as ISO 10218, which set force limits, separation distances, and stopping requirements.
What is a cobot?
A collaborative robot, built with limited force and sensors that stop it on unexpected contact, so it can work next to people instead of behind a cage.
Why do robots struggle with picking up objects?
Grasping requires predicting how an object will behave when touched, which depends on weight, friction, and softness that are hard to sense in advance. Tactile sensing lags far behind vision.
Which country uses the most robots?
China, by a wide margin. It installed roughly 295,000 industrial robots in 2024, more than half of the world total.
Source notes
The origin of the word is documented in the R.U.R. entry, and the first industrial robot in the Unimate record. The first reasoning mobile robot is covered by the Shakey entry, Asimov’s fictional laws by the Three Laws of Robotics entry, and the difficulty of perception and movement by Moravec’s paradox. Robot counts come from the International Federation of Robotics, which reports industrial robots and service robots in separate yearly summaries.
A robot is a machine that couples sensing, computation, and actuation into a loop: it measures something about the world, decides what that measurement means, and moves accordingly. The definition is functional rather than cosmetic, which is why a six-axis arm bolted to a factory floor is a robot and a humanoid shell with no sensing is not. Roughly 4.6 million industrial robots were in operation worldwide in 2024, with about 542,000 installed that year alone.
What the popular picture gets wrong
Three assumptions do most of the damage.
The first is shape. Public imagery is dominated by humanoids, while the articulated arm is the most common industrial robot, with cartesian and gantry units, SCARA arms, and delta mechanisms taking the jobs that suit their geometry. Outside the factory the highest-volume machine is humbler still: close to 20.1 million consumer service robots were sold in 2024, and 97 percent of those were built for domestic tasks such as cleaning floors. Legs are hard to balance and rarely the best answer, and a human form is only advantageous where a robot must operate in spaces built around human bodies.
The second is autonomy. Most deployed robots execute programmed sequences with narrow adaptation. Even surgical robots, which appear sophisticated, are teleoperated: the surgeon commands every motion and the system contributes motion scaling and tremor filtering rather than decisions.
The third is difficulty. Chess fell to computers in 1997. Reliable manipulation of unfamiliar objects has not fallen yet. Hans Moravec described this inversion in 1988: the abilities humans find effortless, perception and movement, are the hardest to automate, while abstract reasoning that humans find taxing is comparatively easy. The usual explanation is evolutionary. Vision and motor control were refined across hundreds of millions of years and run below awareness; symbolic reasoning is recent and deliberate, and deliberate processes are easier to write down as rules.
How an arm actually knows what it is doing
Each joint carries an encoder that reports its angle. A controller compares the measured angle against the commanded angle hundreds or thousands of times per second and drives the motor to close the gap. That feedback loop is the whole basis of industrial accuracy: without it, an arm would drift as loads, wear, and temperature changed.
Turning a desired hand position into joint angles is the inverse kinematics problem, and it is genuinely harder than the forward direction. Forward kinematics, computing where the hand ends up from a given set of joint angles, has exactly one answer and is a direct calculation. Inverse kinematics can have several answers, since many arms can reach the same point with the elbow up or the elbow down, or none at all if the target is outside the workspace.
Certain configurations are worse than merely awkward. At a singularity, typically where two axes align or the arm is fully extended, the arm instantaneously loses the ability to move in some direction, and the joint speeds needed to follow a straight path rise without bound. Motion planners route around these configurations rather than through them.
Geometry is chosen for the job
Six-axis jointed arms are the general-purpose choice. Six is the minimum number of degrees of freedom needed to place a tool at an arbitrary position and orientation, three for position and three for orientation. A seventh axis makes the arm redundant, letting it reach the same pose many ways, which helps with obstacles at the cost of harder planning.
SCARA arms, short for Selective Compliance Assembly Robot Arm, are stiff vertically and slightly compliant horizontally. That combination suits vertical insertion: a part pressed downward can shift sideways enough to self-align into a hole. They typically have four axes and are fast and inexpensive for planar pick-and-place.
Delta robots hang above a conveyor and drive a light platform through three linked arms, with all motors fixed to the frame. Because almost nothing heavy accelerates, they can pick and place several times per second, which is why they appear over food and pharmaceutical packaging lines. The tradeoffs are a small working volume and a payload measured in a few pounds.
Key facts
The word robot comes from Karel Capek’s 1920 play R.U.R.; his brother Josef suggested it, from the Czech robota, forced labor.
Unimate, the first industrial robot, was patented by George Devol (filed 1954, granted 1961) and installed at a General Motors plant in 1961.
Shakey, built at SRI between 1966 and 1972, was the first mobile robot able to decompose a goal into steps by itself.
Isaac Asimov stated the Three Laws of Robotics in the 1942 story Runaround. He coined the word robotics a year earlier, in the May 1941 story Liar!
Six degrees of freedom is the minimum for arbitrary position and orientation.
Industrial arms typically have repeatability an order of magnitude better than absolute accuracy, because accuracy depends on how well the controller’s geometric model matches the real machine.
More than 4.6 million industrial robots were operating worldwide in 2024; China alone installed about 295,000 that year.
Boston Dynamics retired the hydraulic Atlas in April 2024 and replaced it with an all-electric successor, presented for the first time as a product.
NASA’s Ingenuity flew 72 times on Mars between 2021 and 2024, against an original plan of up to five flights.
Autonomy where teleoperation is impossible
Mars makes autonomy mandatory rather than optional. One-way light time between Earth and Mars runs from roughly 4 to 24 minutes depending on orbital positions, so joystick control is physically impossible.
Perseverance addresses this with an onboard navigation system that builds three-dimensional maps of the terrain ahead, identifies hazards, and plans a route around them without new instructions from Earth. Its distinguishing feature is that it evaluates terrain while the wheels are still turning rather than stopping to think, which is what allows meaningfully longer drives per day than earlier rovers achieved.
The same constraint applies less dramatically to underwater vehicles, where radio does not propagate, and to any system that must keep working when a communications link drops.
Safety is a property of the application
Collaborative robots are frequently misunderstood as inherently safe machines. The standards treat collaborative operation as a property of an application, not a certification stamped on an arm.
ISO 10218 covers industrial robot safety, and the technical specification ISO/TS 15066 supplements it with biomechanical limits assigned body region by body region, since a forearm tolerates more contact force than a face. Several distinct modes are defined: safety-rated monitored stop, hand guiding, speed and separation monitoring, which maintains a protective distance that varies with relative speed, and power and force limiting.
The practical consequence is that a force-limited arm holding a blade, a soldering iron, or a hot workpiece is not a safe application. Risk assessment covers the tool, the workpiece, and the layout, not the arm in isolation.
The economics of a robot cell
A common surprise is how little of a robot installation’s cost is the robot. The arm is often a minority of the total, with the balance going to tooling, fixtures, safety equipment, integration engineering, and programming. This is why robot price reductions have moved deployment numbers less than expected, and why a cell built for one part can be expensive to repurpose for another.
It also explains where robots appear first. Industrial installations concentrate in a few high-volume sectors: of the robots installed in 2024, electronics accounted for 24 percent, automotive for 23 percent, and metal and machinery for 16 percent. The International Federation of Robotics points at the surrounding engineering, vision and process design in particular, as what most often prevents adoption at small and medium-sized firms, and at the supply of system integrators who can provide it as the usual bottleneck.
Collaborative robots changed part of that equation. Where a risk assessment permits it, the arm and the operator share a workspace with no perimeter fence, and hand guiding lets an operator record a motion by moving the arm rather than by writing a program. Both remove pieces of the installation that a fenced cell requires before it can run at all.
Common misconceptions
“Asimov’s laws govern real robots.” They are fiction from 1942, and most of the stories turn on the laws producing unintended results. Real safety comes from published standards and physical measures.
“Surgical robots operate autonomously.” They are teleoperated. The system scales hand motion down and filters tremor; the surgeon makes every movement.
“Repeatability and accuracy are the same thing.” Repeatability is returning to the same physical point; accuracy is reaching a point specified in world coordinates. Accuracy is usually the weaker figure and can be improved by calibration.
“Wheel odometry is reliable if the wheels do not slip.” Error accumulates without bound even without gross slip, because tire deformation and surface variation contribute. External references are required to correct it.
“Humanoid robots are the frontier.” Humanoids attract attention, but the hard unsolved problem is general manipulation, which is largely independent of body plan.
Frequently asked questions
Why do most industrial robots have six axes?
Six is the minimum to place a tool at any position and orientation within reach. Fewer restricts the achievable poses; more adds flexibility and planning complexity.
What is a singularity?
A configuration where the arm instantaneously loses the ability to move in some direction and the joint rates needed to follow a path grow without bound. Planners avoid them.
Why can’t Mars rovers be driven live?
Radio takes 4 to 24 minutes each way. Any live command would arrive long after the situation it was meant to address.
What makes a cobot safe?
Force and speed limits, contact sensing, and an application-level risk assessment defined by ISO 10218 and ISO/TS 15066. The arm alone is not the unit of safety.
Why is grasping unfamiliar objects still hard?
It requires predicting contact behavior from properties, including mass distribution, friction, and deformability, that are difficult to sense in advance. Tactile sensing remains far less developed than vision.
Are robots taking manufacturing jobs?
Robot installations have roughly doubled over a decade while manufacturing employment in many countries has not collapsed. The clearer effect is on specific tasks that are dangerous, repetitive, or demand inhuman consistency.
Source notes
Definitions and control basics come from the Robot entry, with the first industrial machine documented in the Unimate record. Kinematics, redundancy, and singularities are covered by robot kinematics, and collaborative safety by the ISO 10218 entry. The relative difficulty of perception and reasoning is described in Moravec’s paradox, mechanical structures in the industrial robot entry, deployment counts from the International Federation of Robotics summaries for industrial robots and service robots, and the humanoid transition from the Atlas entry.
Robotics is the engineering discipline concerned with machines that close a loop between sensing, computation, and physical action. The field’s persistent difficulty is not computation but contact: predicting and controlling what happens when a machine touches a world it can only partially observe. Nearly every hard problem in robotics, from grasping to legged locomotion to assembly, reduces to some form of that difficulty.
Kinematics, redundancy, and the failure modes
Six degrees of freedom is the minimum for placing a rigid body at an arbitrary position and orientation: three for position, three for orientation. A seven-axis arm is kinematically redundant, meaning a given end-effector pose corresponds to a continuum of joint configurations rather than a discrete set. Redundancy is exploited to avoid obstacles, stay within joint limits, and steer clear of singularities, and it converts inverse kinematics from a root-finding problem into an optimization problem.
Singularities are the structural failure mode. At a singular configuration the mapping from joint rates to end-effector velocity loses rank, so the arm cannot move instantaneously in at least one direction, and joint rates required to track a Cartesian path grow without bound as the configuration is approached. Wrist singularities, where two rotational axes align, and boundary singularities at full extension are the common cases. Controllers detect proximity and either damp the commanded motion or refuse it, and planners route paths around these regions.
Accuracy and repeatability are distinct and often conflated. Repeatability describes the spread when the arm returns to the same commanded joint configuration, and is dominated by mechanical hysteresis and encoder resolution. Absolute accuracy describes reaching a pose specified in world coordinates, and is dominated by the mismatch between the controller’s geometric model and the real link lengths, joint offsets, and deflections. A machine quoted at very fine repeatability may be several times worse in absolute accuracy until it is calibrated, which is why offline programming requires calibration and teach-pendant programming does not.
Transmission design
Joint transmissions are where robotics accepts an unusual set of tradeoffs. Strain wave gearing, commonly called harmonic drive, dominates because it delivers reduction ratios from roughly 30:1 to 320:1 in one compact coaxial stage with near-zero backlash, in the volume where planetary gearing would typically manage 10:1. The mechanism deforms a flexible toothed cup with an elliptical wave generator so it engages a rigid circular spline at two regions, engaging many teeth simultaneously.
Backlash matters more in a serial arm than intuition suggests, because angular error at a proximal joint is amplified by the length of every link beyond it. A tenth of a degree of lost motion at a shoulder is about 1.75 milliradians, which over a link roughly a yard (1 m) long displaces the tool by about seven hundredths of an inch (1.75 mm).
Cycloidal drives are the usual alternative where shock loading or higher torque density is required. They apply load through many teeth at once, which gives high torque output for their size, and the price is sliding contact and the vibration an eccentric disc produces unless it is balanced by a second disc or a counterweight.
Compliance and contact
A stiff position-controlled arm is the wrong instrument for contact tasks. Commanding a position against a rigid environment converts any positional error into force through the effective stiffness of the mechanism, and a stiff arm turns a few thousandths of an inch (a fraction of a millimeter) into hundreds of newtons. Assembly, polishing, deburring, and grasping therefore call for regulating the relationship between motion and force rather than position alone.
Impedance control specifies that relationship, so the end effector behaves like a tunable mass, spring, and damper rather than a rigid commanded point. It is distinct from pure force control, which regulates force irrespective of position, and the distinction matters because a pure force controller behaves badly on loss of contact.
Series elastic actuation places a deliberate compliant element between the geared motor and the load. Deflection of that element is proportional to the transmitted force, so a position measurement across the spring becomes a force measurement, and the same spring softens an accidental collision instead of passing it rigidly through the drive. The cost is control bandwidth: the spring introduces oscillatory dynamics that limit how quickly force can be commanded and degrade precise position tracking, which is the reason not every robot uses them.
State estimation
Dead reckoning from wheel encoders accumulates unbounded error. Slip, tire deformation, and surface variation contribute error that no improvement in encoder resolution addresses, because the drift is a property of the method rather than the sensor. Any robot navigating to a goal over distance requires external references.
Simultaneous localization and mapping addresses the circular dependency directly: position estimation requires a map, and mapping requires known position, so both are estimated jointly. The step that keeps the estimate from degrading is loop closure, recognizing a previously visited place and redistributing accumulated drift across the trajectory that connects the two observations. Without loop closure, a SLAM map bends progressively as the robot travels, and corridors that should meet do not.
Learning and the reality gap
Policies trained in simulation frequently fail on hardware. Physics engines approximate contact and friction, and they omit actuator latency, sensor noise, and mechanical compliance that the real system exhibits.
Domain randomization addresses this by varying simulation parameters across training episodes, including masses, friction coefficients, control latencies, and visual appearance, so no learned policy can depend on any single parameterization. If the policy holds up across a wide distribution in simulation, reality becomes one more draw from that distribution rather than an unmodeled case.
Contact remains the dominant contributor to the gap. Rendering fidelity has improved considerably faster than contact and deformation modeling, which is why vision-based policies transfer more readily than contact-rich manipulation policies.
Why manipulation resists
Moravec’s observation, stated in 1988, holds with unusual durability. Chess fell in 1997. General-purpose grasping has not.
The specific obstacle is not gripper force but state estimation at the contact. Predicting how an object responds to being touched requires knowledge of mass distribution, surface friction, and deformability, none of which is directly observable before contact and only partially observable during it. Tactile sensing lags vision substantially in resolution, durability, and available data, so the gripper typically has poor information about the very interaction it is trying to control.
This is also why unfamiliar objects and unstructured settings remain the dominant open failure mode. Generalist manipulation policies now reach roughly the same task performance in homes they were never trained on as in ones they were, yet the residual failures cluster exactly where contact state is hardest to infer: an unfamiliar drawer handle, or a spill the arm’s own body is occluding. No system does these tasks as reliably as a person.
Safety as a system property
ISO 10218 covers industrial robot safety requirements, and the technical specification ISO/TS 15066, since folded into an updated edition of the standard, supplements it with biomechanical limits assigned per body region, reflecting differing tolerance between a forearm and a face.
Four collaborative modes are defined: safety-rated monitored stop, hand guiding, speed and separation monitoring, which maintains a protective separation distance calculated from relative speed and stopping performance, and power and force limiting.
The conceptual point is that collaborative operation is a property of an application, not a certification carried by a machine. A force-limited arm holding a blade or a hot workpiece constitutes an unsafe application regardless of the arm’s rating. Risk assessment covers end effector, workpiece, trajectory, and layout.
Actuation choice
Hydraulic actuation offers very high force density and tolerates impact loading well, which is why it dominated early dynamic legged robots. Its costs are a power unit, fluid handling, noise, leakage, and maintenance burden that make it unattractive outside research. Boston Dynamics retired its hydraulic Atlas in April 2024 and replaced it with an all-electric successor, presented for the first time as a product rather than a research platform. That transition tracks a general shift: electric actuation with high-ratio gearing or quasi-direct drive has closed enough of the force-density gap to displace hydraulics in most applications.
The gearing decision follows from the control objective. High reduction ratios give torque and holding capability but introduce reflected inertia and friction that make output force hard to sense or regulate, which is acceptable for position-controlled industrial arms and poor for contact-rich legged locomotion. Quasi-direct drive uses low reduction so the motor’s torque is felt at the output almost directly, giving good force transparency and impact tolerance at the price of torque density and thermal headroom.
Programming methods
Two paradigms coexist in industry, and the difference is largely about whether absolute accuracy matters.
Teach-pendant programming moves the arm to each required pose by hand and records joint configurations. Because the recorded values are joint angles rather than world coordinates, the method depends only on repeatability, and no calibration is required. It remains the dominant method in practice and is well suited to a workcell that will run one task for years.
Offline programming generates trajectories from a CAD model of the cell, in world coordinates. This requires the controller’s geometric model to correspond to the physical machine, so it depends on absolute accuracy and therefore on calibration. It pays off where cells are reconfigured often, where the part geometry comes from CAD anyway, or where the robot must not stop production to be reprogrammed.
Key facts
Six degrees of freedom is the minimum for arbitrary pose; a seventh introduces a continuum of solutions.
At a singularity, the joint-rate-to-end-effector-velocity mapping loses rank and required joint rates diverge.
Repeatability is typically an order of magnitude better than uncalibrated absolute accuracy.
Strain wave gearing gives roughly 30:1 to 320:1 reduction in one compact stage with near-zero backlash.
Impedance control regulates the relationship between displacement and force; series elastic actuators make force directly measurable at the cost of bandwidth.
Odometry drift is unbounded and is corrected by external references; SLAM loop closure redistributes accumulated error.
Domain randomization narrows the simulation-to-reality gap; contact dynamics remain its largest component.
ISO 10218 and ISO/TS 15066 define four collaborative modes and per-body-region force limits.
Common misconceptions at expert level
“Redundancy eliminates singularities.” It allows an arm to select configurations that avoid many singular poses. The singularities remain properties of the mechanism.
“Better encoders fix odometry drift.” Drift arises from slip and contact geometry, not measurement resolution. Only external references bound it.
“Impedance control and force control are interchangeable.” Impedance regulates a displacement-force relationship; pure force control does not, and behaves poorly on contact loss.
“Sim-to-real is a rendering problem.” Contact and friction modeling dominate the gap. Visual fidelity has advanced faster than contact fidelity.
“A collaborative-rated arm makes an application safe.” Safety is assessed for the application, including tooling and workpiece.
Frequently asked questions
Why is inverse kinematics harder than forward kinematics?
Forward kinematics is a direct composition of transforms with one answer. Inverse kinematics may admit multiple solutions, a continuum for redundant arms, or none, and closed-form solutions exist only for particular geometries.
What actually limits absolute accuracy?
Mismatch between the controller’s kinematic model and the physical machine, including link length tolerances, joint offsets, and load-dependent deflection. Calibration identifies and compensates these.
Why do legged robots use series elastic or quasi-direct-drive actuation?
Both improve tolerance to impact and enable force control at the foot. Highly geared stiff joints transmit impact into the gearbox and make ground contact force hard to regulate.
What makes loop closure difficult?
Recognizing a place already visited despite different viewpoint, lighting, or partial occlusion, while avoiding false positives. A wrong loop closure corrupts the map more severely than accumulating drift.
Why do manipulation policies generalize worse than navigation policies?
Navigation operates largely on geometry that sensors observe directly. Manipulation depends on physical properties revealed only through contact, and errors are unforgiving because contact forces rise steeply.
Is the humanoid form technically justified?
Chiefly where a robot must operate in environments and with tools designed around human bodies. Where the environment can instead be built around the machine, geometry is chosen for the task, which is why cartesian gantries, SCARA arms, and delta mechanisms all persist alongside the jointed arm.
