A robot welds a car body to a tenth of a millimetre all day. Ask it to pick a sock out of a basket and it cannot

An industrial robot will weld a car body to a tenth of a millimetre, eight hours a day, for a decade, without a single deviation. The same robot, presented with a basket of laundry and asked to pick out one sock, is helpless. So is almost every robot built.

This inversion is not a temporary engineering gap. It is a structural feature of the field, and it has a name — Moravec's paradox: the things humans find difficult, such as arithmetic and chess, turn out to be easy for machines, while the things humans do without any conscious effort at all, such as seeing, walking and picking things up, turn out to be extraordinarily hard.

The usual explanation is evolutionary. Sensorimotor skill has been under optimisation for hundreds of millions of years; abstract reasoning is a recent and thin layer. The oldest capabilities run beneath awareness and feel effortless, which is exactly why we badly underestimate what they involve.

The factory is built around the robot, not the other way round

The first thing to understand about a working industrial robot is how much of its competence lives outside the robot.

A welding cell works because the car body arrives in precisely the same position every time, held by a jig that removes all uncertainty. Parts are presented in fixtures. Lighting is controlled. The sequence is identical on every cycle. Under those conditions the robot does not need to perceive anything — it needs to repeat a trajectory accurately, which is a solved mechanical problem.

Remove the structure and performance collapses. The robot has no notion of what it is doing; it has a path. This is why industrial automation succeeded first in high-volume manufacturing of identical items, and why the same technology has made so little progress in a kitchen, a repair workshop or a field.

The distinction engineers draw is between structured and unstructured environments. Enormous amounts of industrial engineering consist of converting the second into the first — building jigs, feeders and conveyors so that the robot never encounters a surprise. Where you cannot afford to restructure the world, the robot has to cope with it, and that is where the field runs into its real problem.

Grasping is a sensing problem, not a gripping problem

Picking something up sounds like the simplest possible task. Consider what it actually requires.

You must know where the object is and how it is oriented — which means perceiving it in three dimensions, often partly hidden behind other objects. You must infer what it is made of, because a glass tumbler, a ripe tomato and a steel bolt need entirely different amounts of force. You must estimate where its centre of mass is, or it will rotate out of your grip as you lift. You must detect slip the instant it begins and tighten before the object falls. And you must do all of this with an object you may never have seen before.

Humans do this using a dense array of touch sensors — the hand contains many thousands of mechanoreceptors reporting pressure, vibration and slip continuously — combined with a hand of roughly two dozen independently controllable degrees of freedom. A typical industrial gripper has one degree of freedom, opening and closing, and no tactile sensing whatsoever. It knows position, and possibly gripping force, and nothing else.

That is the gap. Robot vision has improved enormously; robot touch has not, and touch is what grasping runs on. A great deal of current research is on tactile skins, optical tactile sensors that watch a deformable pad from the inside, and grippers that succeed without precise sensing — soft and underactuated hands that conform passively to whatever they close around, suction cups, and the jamming gripper, a bag of granular material that moulds to an object and then stiffens when air is pumped out. These work by removing the need for accurate perception rather than by supplying it.

Industrial robots do not understand the objects they handle. They repeat a motion, and the factory is engineered so that repeating that motion is enough.

Why contact breaks a robot's control system

There is a deeper reason manipulation is hard, and it sits in the control system.

An industrial robot is usually a position-controlled machine: it is commanded to place its end effector at a coordinate, and it is stiff enough to get there against considerable resistance. That stiffness is what makes it accurate. It also means that if the commanded position is one millimetre inside a table, the robot will attempt to move one millimetre into the table, with whatever force it takes.

Almost every interesting manipulation task involves contact — inserting a peg, wiping a surface, closing a fastener, handing something to a person. Contact makes pure position control dangerous and often impossible, because the exact position of the contact is never known to the accuracy the robot is capable of. The answer is compliance: force control and impedance control, where the robot is commanded to behave as though it had a certain springiness, pushing with a specified force rather than reaching a specified point. A robot doing an assembly task well looks slightly loose, and that looseness is the control strategy.

This also explains a persistent difficulty with machine learning approaches. Policies trained in simulation transfer poorly to real manipulation, because the physics that matter most at contact — friction, deformation, micro-slip, the behaviour of a cable or a piece of cloth — are precisely the things simulators model worst. Locomotion and navigation have crossed the simulation-to-reality gap far more successfully than manipulation has, and contact physics is the reason.

Where the economics actually bite

There is an Indian dimension to this that is usually discussed badly, in terms of robots replacing workers.

The relevant variables are volume, variety and wages. A robot is a large fixed cost that pays back through repetition, so automation is most attractive where the same operation is performed millions of times, quality requirements are strict, and labour is expensive. That describes automotive manufacturing, which is where Indian industrial robot deployment is concentrated, and it describes very little else in the country's manufacturing base.

Much Indian manufacturing runs at low volume with high product variety — short runs, frequent changeovers, rapidly changing specifications. That is the exact regime in which conventional robots are weakest, because every change requires reprogramming and refixturing, and the fixed cost is amortised over far fewer units. Meanwhile, a skilled worker handles variety effortlessly, which is precisely the capability robots lack.

The technologies that could change that equation are the ones aimed at reducing setup cost rather than increasing speed: collaborative robots that work safely alongside people without cages and can be moved between tasks, teaching by demonstration rather than by programming, and vision systems that tolerate parts arriving in approximate positions. Whether those mature into something economic at Indian volumes is an open question, and a more useful one than the replacement debate.

What is genuinely changing

Three things are moving. Learning-based grasping has improved substantially — systems trained on very large numbers of grasp attempts now generalise to unfamiliar objects far better than hand-coded approaches ever did. Cheaper force and torque sensing is making compliance available on less expensive hardware. And the newer approach of using large vision and language models to interpret a task instruction and decompose it into steps has made robots much better at deciding what to do, though the physical execution remains the bottleneck.

It is worth being precise about what has and has not been solved. Navigation is largely solved — warehouse robots move shelves reliably. Perception has improved dramatically. What remains stubborn is the last stage: the contact between a machine and an object whose properties it must infer. In warehouse automation this is explicitly called the picking problem, and it is why the moving of goods has been automated far ahead of the handling of them.

Why it matters for students and researchers

Robotics is one of the few engineering fields where mechanical design, control theory, sensing and computation cannot be separated — a better gripper can substitute for a better algorithm, and a compliant mechanism can solve in hardware what would be a difficult control problem in software. That makes it unusually rewarding to work on and unusually easy to approach from several directions.

The accessible research problems are real. Low-cost tactile sensing, where much of the recent progress has come from cameras watching deformable materials rather than from expensive sensor arrays. Underactuated and compliant gripper design, which is mechanism design and can be prototyped with 3D printing. Automation for small and medium manufacturing, where the binding constraint is changeover cost rather than cycle time and where the Indian industrial context poses the question sharply. Safety and interaction in shared workspaces. And agricultural robotics, where the objects are deformable, variable, occluded by leaves and easily damaged — quite possibly the hardest manipulation environment there is, and one with obvious local relevance.

That breadth — industrial, factory and building automation, robotics and related automation technologies — is the stated scope of the International Journal of Advanced Robotics and Automation Technology, a peer-reviewed journal launched in 2023. For mechanical engineering students, the useful correction is this: robots are not held back by a shortage of intelligence. They are held back by hands, by touch, and by the difficulty of acting on a world that will not hold still.

Frequently asked questions

What is Moravec's paradox?

The observation that tasks humans find hard, such as arithmetic and chess, are comparatively easy for machines, while tasks humans perform effortlessly, such as perception and manipulation, are extremely difficult. Sensorimotor ability has a far longer evolutionary history and runs below conscious awareness, so we underestimate it.

Why can industrial robots be so precise and still not pick up a sock?

Because their precision comes from repeating a fixed trajectory in an environment engineered to remove uncertainty. They do not perceive or understand the objects involved, so when the object's position, shape or material varies, the approach fails.

What makes grasping so difficult for a robot?

It requires knowing an object's position and orientation, inferring its material and weight distribution, applying appropriate force, and detecting slip immediately. Humans do this with thousands of touch receptors and a highly articulated hand; most robot grippers have one degree of freedom and no tactile sensing.

Why does a robot need compliance to do assembly?

Because a stiff position-controlled robot will drive into whatever it contacts if the commanded position is even slightly wrong. Force and impedance control let the robot push with a specified force and yield on contact, which is what makes tasks such as inserting a part possible.

Why do robots trained in simulation fail at real manipulation?

Because contact physics — friction, deformation, slip, and the behaviour of flexible materials — is the part of the world that simulators reproduce least accurately. Navigation transfers from simulation far more successfully than manipulation does.

Where does automation make economic sense?

Where the same operation repeats at high volume, quality requirements are strict and labour is expensive. Low-volume, high-variety production is the regime where conventional robots are weakest, because changeover and reprogramming costs are spread over far fewer units.