A robot’s grip is not a single capability. Picking up a glass, folding a shirt, and opening a drawer may look like variations of the same simple idea: move a hand, make contact, apply force. Physically, they are very different problems. A glass can slip, crack, or spill. Fabric has no stable shape and can conceal its own edges. A drawer moves only along a constrained path and may resist for reasons the robot cannot see.

This is why robot manipulation remains difficult even as robots become better at recognizing images, following spoken instructions, and navigating familiar spaces. The hard part is not merely deciding what to do. It is acting through contact with a world full of uncertainty: different materials, hidden weight, changing friction, flexible surfaces, clutter, wear, and motion that is often impossible to predict perfectly.

Reliable manipulation sits at the meeting point of perception, mechanics, control, and prediction. A robot must see enough to approach an object, feel enough to know what happened at contact, and control its body well enough to adjust before a minor error becomes a dropped glass, a tangled cloth, or a jammed drawer.

Contact is a physics problem before it is a software problem

Every grasp begins with forces. When a robot finger presses against an object, it produces a normal force: the force directed into the surface. That pressure creates the conditions for friction, the resistance that helps prevent one surface from sliding across another.

Friction is essential, but it is not guaranteed. A dry rubber gripper and a smooth carton may hold well; the same gripper on a wet, dusty, oily, or polished object may slip at a much lower force. The robot therefore needs to apply enough inward force to keep an object from falling while avoiding so much force that it dents packaging, crushes food, damages a fragile item, or pushes the object out of position.

Geometry matters as much as force. Where the fingers touch determines whether a grasp can resist twisting and tipping. If a robot lifts an object away from its center of mass—the effective balance point of its weight—the object creates a turning effect called torque. A bottle held near its cap, for example, may rotate downward unless the hand can counter that torque.

Pressure is also distributed, not abstract. The same total gripping force can be safe when spread across broad, compliant finger pads and damaging when concentrated on a small edge. This is one reason a useful robot grasp is more than a command to close a gripper. It is a continuously managed physical relationship between hand, object, gravity, and surrounding surfaces.

Stability depends on what can go wrong

A stable grasp is one that can resist expected disturbances: gravity, acceleration, vibration, a change in orientation, or contact with another object. But the expected disturbance depends on the task. Carrying a sealed can across a table is not the same as tilting a cup, pulling a cable, or inserting a plug.

For this reason, robot grasping systems often need to choose not only whether they can pick something up, but how and where to hold it. A grasp that succeeds at lifting may be poor for pouring. A grip that is stable for carrying may block the handle the robot needs to turn.

A glass demands precision, feedback, and restraint

Consider an ordinary drinking glass. Its smooth surface can offer limited friction. Its walls may be thin. If it contains liquid, the robot is handling not one rigid body but a container and a moving fluid. A fast acceleration or abrupt stop can cause the liquid to slosh, shifting the combined center of mass and increasing the chance of a spill.

Before contact, robotics perception can help estimate the glass’s location, orientation, size, and accessible sides. Cameras provide visual detail, while depth sensors can help infer three-dimensional shape and distance. Yet a clear glass exposes a recurring weakness of machine perception: transparent and reflective surfaces can confuse systems that rely on visible texture or projected depth patterns. Reflections, condensation, glare, and an unknown liquid level add uncertainty.

After contact, sensing becomes especially important. A robot may infer that its fingers have touched the glass from joint motion, motor effort, fingertip force readings, or tactile sensors. It can then close gradually and monitor whether contact forces are building as expected. If the glass shifts, the robot may need to reposition its fingers or alter its grip rather than continue a preplanned motion.

This is the practical meaning of force control in robotics. Instead of commanding only a finger position—close by a specified amount—the system manages the force exchanged with the object. Position is still important, but a rigidly executed position command can be a bad fit when the glass is slightly wider than expected, tilted, fragile, or not quite where vision predicted.

A good robotic grasp of a glass is therefore cautious and adaptive. It does not assume that the visual model is complete. It treats contact as new evidence.

Fabric behaves like a different physical world

Rigid objects are challenging because they can be slippery, cluttered, or irregular. Fabric introduces a more fundamental problem: it changes shape continuously. A towel does not have one fixed geometry that a robot can identify, grasp, and preserve. Its folds create new surfaces, hide corners, and change frictional contact from moment to moment.

For a person, finding the corner of a shirt or separating two layers of cloth is familiar work. For a robot, it can be ambiguous. A visible edge may be a hem, a crease, an overlap, or a shadow. A pinched region may include one layer or several. Pulling one point may flatten the fabric, tighten it, bunch it, drag it across a table, or simply cause the grasp to fail.

That makes dexterous manipulation of fabric an exercise in active sensing. The robot may need to pull gently to reveal tension, regrasp after a fold shifts, or use a table as a constraint that limits how the material can move. Rather than trying to solve everything in a single image, it can perform small actions that make the state of the cloth easier to infer.

Small contacts can create large changes

Deformable materials are difficult to predict because their response is distributed. Pressing or pulling at one point changes the shape elsewhere. The outcome depends on material thickness, weave, stretch, wrinkles, friction against the robot and work surface, and whether parts of the fabric are caught underneath other parts.

Physics-based models can represent some of these effects, but real textiles vary widely and are hard to characterize perfectly. Learned systems can recognize patterns from examples, but may struggle when a new material, lighting condition, or arrangement falls outside their experience. In practice, robust fabric handling is likely to rely on both: useful prior models and frequent feedback-driven correction.

This is why folding fabric is not simply “grasping a soft object.” The robot must reason about edges, layers, tension, and changing shape over time.

Opening a drawer is an interaction, not a pickup

A drawer presents another kind of manipulation problem. The goal is not to lift an object freely through space. It is to move an articulated mechanism along a path imposed by rails, slides, or runners. The robot must locate a handle or usable surface, make a stable connection, pull in the correct direction, and recognize whether the drawer has actually started moving.

Resistance can come from several places: friction in the slides, the drawer’s load, a latch, poor alignment, or contact with an obstacle. Pulling harder is not always the correct response. If the robot’s hand is misaligned, extra force may twist the handle, increase sideways loading, destabilize the grasp, or damage a component.

The useful behavior is often to yield slightly while maintaining intent. This is where compliant robots and compliant control matter. Compliance means controlled give: the robot does not behave as though every position must be enforced with unyielding stiffness. It can accommodate small errors in pose and contact while still applying a purposeful pull.

In engineering terms, impedance control shapes the relationship between motion and force, often making the robot behave as if it has chosen levels of stiffness and damping. Admittance control takes a related approach from the other direction, using measured forces to adjust motion commands. The exact implementation varies by robot and task, but the shared aim is to make contact safer and more tolerant of uncertainty.

A drawer makes the central lesson visible: motion and force are coupled. The robot cannot understand the task by planning a path alone. It must feel whether the physical world is allowing that path.

The sensing stack: seeing is not feeling

No single sensor gives a robot a complete account of contact. A capable manipulation system may combine several sources of information:

• Cameras can identify objects, handles, surfaces, and scene context, but can be blocked by the robot’s own hand or confused by glare and occlusion.
• Depth sensors can help estimate shape and distance, although performance can vary with surface material, lighting, range, and sensor design.
• Joint-position sensors report where robot links and fingers are, but not necessarily what they are touching.
• Motor-current estimates and joint-torque sensing can indicate resistance or unexpected contact, though they may be indirect and affected by friction or transmission mechanics inside the robot.
• Fingertip force sensors can measure contact loads at specific points.
• Tactile arrays can provide richer contact patterns, potentially revealing pressure distribution, contact location, and signs of slip.

Robot tactile sensing is particularly valuable once a hand reaches an object. Vision may say a finger should be on the side of a cup; touch can reveal whether it is actually on the cup, pressing too hard, resting on a label seam, or missing altogether.

Combining these signals is difficult. Measurements arrive with different delays, levels of noise, and coordinate systems. The robot must know where its own hand is, how its sensors are aligned, and how to interpret a force reading while the arm is moving. Sensor fusion is not merely adding data together; it is building a consistent estimate of an evolving physical situation.

Position control, force control, and controlled give

Position control tells a robot to move toward a target pose: place the hand here, rotate the wrist there, close the fingers this far. It can be highly accurate in structured settings, especially when objects are consistently located and fixtures constrain the work.

But position control alone can be brittle around uncertain contact. If a robot expects a surface to be two millimeters farther away than it is, a stiff motion may create an unexpectedly large force. If it expects a handle in exactly one location, it may miss or collide when the drawer is slightly open or shifted.

Force control makes the desired interaction explicit: press lightly, maintain a certain contact load, pull until resistance changes, or stop when force rises beyond a safe threshold. Compliance complements this by allowing the robot to absorb small discrepancies rather than fighting them.

There is always a trade-off. Too much stiffness can make a robot precise but unforgiving. Too much compliance can make it safe and adaptable but less accurate, especially for tasks requiring firm alignment. Successful systems choose the balance that fits the object, tool, environment, and consequences of error.

Manipulation requires prediction—and continual revision

Before a robot acts, it needs some prediction of what will happen next. Will the object slide if pushed? Will it rotate when lifted? Is the lid flexible, the package crushable, the door latched, or the cable snagged? These questions depend on properties that are often only partly known: mass, mass distribution, friction, stiffness, elasticity, articulation, and deformability.

Physics-based robotics uses models of mechanics to plan or evaluate actions. Such models can be especially useful when a system knows the relevant geometry and physical parameters. They also face a hard reality: contact is messy. Small differences in surface finish, wear, moisture, object contents, or collision timing can change the result.

Learning-based approaches offer a different route. A robot can learn policies from demonstrations, trial and error, or large collections of recorded interactions. These methods can capture patterns that are difficult to write down explicitly. Yet learning does not remove physics. A learned policy still has to cope with real forces, imperfect sensing, and unfamiliar cases.

Hybrid approaches are increasingly natural: use visual and learned representations to identify promising actions, then use physical models and rapid feedback loops to execute them safely. The key is that prediction cannot be a one-time calculation. Every touch produces new information, and the robot must update its plan while acting.

Why a general-purpose robot hand is still hard to build

The household world has a long tail of variation. Containers differ in shape, material, seal, weight, and contents. Handles are worn, hidden, soft, loose, or placed near obstacles. Objects may be wet, transparent, dented, stacked, or partly covered. Lighting changes. Floors and counters have different friction. A task that looks routine to a person can contain many unspoken assumptions.

A robot can perform impressively in a controlled demonstration and still struggle when those assumptions change. Simulation helps researchers test many possible scenes, but simulated contacts cannot capture every real material variation, sensor imperfection, calibration error, or delay in the control loop. Bridging this sim-to-real gap remains a central engineering challenge.

Better language models and better visual recognition can make robots more capable at interpreting instructions and scenes. They do not automatically provide robot hand-eye coordination at the point of contact. A system may correctly identify a mug and still fail to grasp it if it misjudges the rim, overlooks a wet surface, or cannot regulate force quickly enough.

Nor is a humanlike hand always the best answer. Task-specific end effectors, suction tools, parallel-jaw grippers, soft grippers, and adaptive fingers can be more reliable for particular jobs. The most effective design is often the one that reduces uncertainty rather than attempting to imitate every feature of a human hand.

Reliable manipulation will include recovery, not just success

The future of robot manipulation is unlikely to be defined by a machine that never makes a mistake. More useful systems will recognize uncertainty, slow down around fragile objects, choose safer grasps, use tables and fixtures intelligently, and recover when the first attempt fails.

That may mean reorienting an object after a partial grasp, changing grip when slip is detected, backing away after unexpected resistance, or asking a person for assistance when confidence is low. In logistics, manufacturing, healthcare, homes, and space operations, this kind of graceful failure can matter as much as peak dexterity.

The glass, the fabric, and the drawer are three reminders that physical intelligence is local. It happens at the fingertip, at the edge of a fold, and at the moment a handle begins to move. Progress will come not from treating manipulation as a software feature, but from building machines that can perceive, predict, and adapt to the real mechanics of contact.

Image by DimaDim_art on Pixabay.