When You Don't Need to Grasp
Most of this track has been about closing a hand around an object — a grasp that pins the object so firmly it cannot escape your fingers. But think about how you actually move things in the real world. You slide a coffee mug across the table toward you. You nudge a fallen pen until it pokes past the edge where you can finally pick it up. You tip a heavy box onto one corner to walk it into place. In none of these did you wrap your hand around the object first. This whole family of moves — pushing, sliding, tipping, toppling, rolling — is called non-prehensile manipulation: changing where an object is without ever holding it.
Why bother, when grasping seems more direct? Sometimes the object is too big, too flat, or too flush against a wall for any gripper to fit around it — a coin on a table or a book wedged on a shelf has nowhere for fingers to go. Sometimes pushing is simply faster: shoving a cluttered pile aside beats picking up each item one by one. And sometimes a pre-push sets up an easier grasp later — sliding an object to the table's edge so the fingers finally have clearance underneath. Non-prehensile moves turn the environment — table, wall, gravity — into an extra helping hand.
The catch is that pushing is surprisingly hard to predict. When you grasp an object, it becomes rigidly attached to your hand and moves exactly as the hand moves. When you push it, the object can rotate, slip, or skid in ways that depend on friction you cannot see and a mass distribution you can only guess at. Push a long ruler near one end and it pivots; push it at the center and it slides straight. The robot has to reason about contact dynamics — how forces at the contact point translate into the object's motion — and that reasoning is the heart of this final chapter.
Staying Soft: Contact-Rich Tasks and Compliance
Now consider a different challenge: inserting a key into a lock, seating a peg in a hole, plugging in a USB connector, or wiping a sponge along a curved surface. These are contact-rich manipulation tasks — the whole point of the motion is to keep touching something, often along an edge or inside a tight clearance. Here the danger is the opposite of pushing's unpredictability. The robot knows roughly where things are, but if it pushes a fraction of a millimeter too far, it jams, scratches, or snaps the part.
A traditional position-controlled robot is dangerous here because it is stubborn. Tell it to go to a point and it will drive there with enormous force, ignoring anything in the way. If the hole is one millimeter off from where the robot thinks it is, the robot does not gently feel for it — it slams the peg into the surface and keeps pushing, because its only goal is to reach the commanded position. What we want instead is a robot that yields on contact, the way your own arm does when you feel resistance.
The elegant answer is impedance control. Instead of commanding a position and demanding the robot reach it no matter what, you command a relationship between force and motion — you make the robot behave like a spring. Push against an impedance-controlled arm and it pushes back gently, in proportion to how far you have displaced it, then springs back when you let go. You can dial the stiffness: stiff like a steel rod when you need precision, soft like a sponge when you need to absorb contact. The closely related admittance control reaches a similar softness from the other direction — it measures the contact force and moves in response. Either way, the robot stops fighting the world and starts negotiating with it.
Softness can also be built into the hardware, not just the software. A remote center of compliance is a clever passive mount near the gripper that lets the held part tilt and shift slightly under contact forces, so a slightly misaligned peg self-corrects as it slides into the hole — no sensor or computation required. And a series-elastic actuator deliberately puts a spring between the motor and the joint, so the limb is forgiving by construction. Good contact-rich manipulation usually blends both: compliant hardware for fast, instinctive give, and impedance control for the precise, programmable feel.
Coping When the World Is Uncertain
Step back and notice the common thread. Pushing is hard because friction is hidden; contact tasks are hard because a millimeter of error can be catastrophic. Both are really faces of one deeper problem — manipulation under uncertainty. The robot never knows the world exactly. Its camera reports the object is here, but its depth estimate is off by centimeters. Its model says the surface is dry, but it is faintly oily and the object slips farther than expected. Its fingers close on the cup, but the cup is heavier on one side and rotates as it lifts.
The first defense is feedback — and not just from the eyes. A force/torque sensor at the wrist or tactile sensors on the fingertips let the robot feel contact as it happens, far faster and more reliably than vision can. When you insert a key in the dark, you do not stare at the lock; you feel the key catch the edge of the slot and wiggle it until it drops in. A robot does the same: it presses gently, senses the side force telling it the peg is off-center, and slides toward the lower resistance until the part seats. This sensing closes the loop tightly enough that the robot recovers from errors it never explicitly saw.
The second defense is to reason about uncertainty rather than ignore it. Instead of betting everything on a single best guess of where the object is, the robot can carry a whole cloud of possibilities — a probability distribution — and act in a way that works for most of them. This is the same probabilistic thinking that powers a particle filter in localization, now turned toward grasping. A famous trick is the funnel: design the motion so that a wide range of starting errors all get squeezed into the same correct final state. A chamfered hole funnels a misaligned peg to the center; tilting and tapping a part against a corner funnels it into a known orientation. You do not need to know exactly where the object started — you only need a plan that converges.
loop until seated:
move down a little under impedance control # stay soft
read wrist force F # feel the contact
if F sideways is large: # peg is off-center
slide toward the direction of least resistance
if F downward drops suddenly: # peg fell into the hole
doneThe Learning Frontier and Where Manipulation Is Heading
Everything so far assumed we could write down the physics — the friction, the geometry, the contact model. But contact is notoriously messy: a thousand tiny micro-collisions and slips that no equation captures cleanly. This is why much of the modern frontier turns to learning. Rather than hand-coding how a push will play out, the robot watches many examples and learns a control policy — a mapping from what it senses to what it should do next.
Where do the examples come from? Often from people. In imitation learning, a human teleoperates the robot through a tricky insertion or a delicate fold dozens of times, and the robot learns to copy that behavior. Or the robot practices in a fast physics simulator, attempting millions of pushes it could never afford to try on real hardware, then transfers what it learned to the physical world — a leap called sim-to-real transfer. The stubborn obstacle is that simulated friction and contact never perfectly match reality — the famous reality gap — so what works in simulation can stumble on a real, slightly slippery table.
The newest and most ambitious direction tries to give robots general competence rather than one trick at a time. A vision-language-action model is trained on huge amounts of robot experience so it can take a camera image and a plain-language instruction — 'put the red block in the bowl' — and output the motions directly, even for objects and phrasings it never saw in training. These models hint at a future where you describe a manipulation task in words and a robot simply does it, the way a capable person would.
Yet the oldest lesson still holds. Grasping a cup, pushing a box, plugging in a cable — the things a toddler does without thinking — remain among the hardest problems in robotics, a stubbornness so famous it has a name, Moravec's Paradox: what is effortless for humans is brutally hard for machines, and what is hard for us — arithmetic, chess — is easy for them. The frontier of manipulation is not a single breakthrough but a steady marriage of all the ideas in this track — careful grasp planning, compliant contact, honest reasoning about uncertainty, and learning from experience — slowly closing the gap between a robot's hand and a human's.