Where should I grab? Grasp planning as a search
Picking something up feels effortless when you do it, but a robot has to answer a precise question first: out of all the ways my hand could touch this object, which one will actually hold it? Choosing that answer is grasp planning — a search over candidate hand placements, each scored and ranked until the best one wins. A grasp here simply means a set of contacts between the hand and the object that, taken together, keep it from slipping away.
But you cannot search for a grasp on an object you cannot locate. So planning almost always starts after object pose estimation — the perception step that works out where the object sits and how it is turned, usually as a position plus an orientation in the camera's view. Feed the planner a confident pose and it can reason about real surfaces; feed it a shaky one and even a clever grasp may close on empty air.
The antipodal grasp: two opposing contacts
The most common grasp a planner proposes is the antipodal grasp: two contact points that face each other across the object, ideally on roughly parallel surfaces, so the line joining them runs nearly straight into both surfaces. Think of how you pinch a book between thumb and forefinger — the two pads press inward along the same line, from opposite sides. Squeeze, and the object has nowhere to slip.
Why does this work so reliably? Because if the two contacts can generate enough friction, squeezing harder fights any twist or pull the object tries to make — a property called force closure. A grasp has force closure when its contacts, using friction, can resist a push or twist from any direction. Its stricter cousin, form closure, cages the object so completely with geometry alone that it cannot move even if the surfaces were frictionless. Most two-fingered grasps rely on force closure; form closure usually needs more fingers or a wrapping shape.
Three families of gripper, three kinds of object
The grasp the planner chooses depends heavily on what hand is attached to the robot. That hand is the gripper, and three families dominate, each best suited to a different kind of object.
- Two-jaw parallel grippers — two flat fingers that pinch together. Simple, strong, cheap, and a perfect match for the antipodal grasp. Great for boxes, bottles, and parts with a pair of graspable parallel faces; poor at delicate, irregular, or floppy items.
- Multi-finger hands — three or more fingers, often jointed, that can wrap around a shape. They can achieve form closure and even regrip an object after picking it up. This opens the door to dexterous manipulation and in-hand manipulation, where the hand reorients the object using its own fingers. The cost is complexity: more motors, harder control, and more that can go wrong.
- Suction grippers — a cup that seals against a surface and pulls a vacuum behind it. A suction gripper needs only one smooth, fairly flat, airtight patch, not two opposing faces, so it shines on flat boxes, glossy bags, and sheets that a pincer would struggle with. It fails on porous, sharply curved, or punctured surfaces, where the seal leaks.
Notice how the hardware changes the question. A parallel jaw asks the planner to find two opposing faces. A suction cup asks only for one good sealing patch. A multi-finger hand asks the richest question of all — where should each finger be placed? — which is exactly why it is powerful and why it is hard.
From candidate to contact: the pick pipeline
Once the planner has proposed several grasps, the robot ranks them and executes the winner. Two ideas glue this together. First, each candidate gets a grasp quality metric — a single number estimating how stable and robust the grasp is, so the planner can sort the good from the risky. Second, the arm does not lunge straight at the object; it follows a pre-grasp and approach, lining the hand up a short distance away and then moving in cleanly along the approach direction.
- Estimate the object's pose and propose a handful of candidate grasps on its surfaces.
- Score each candidate with a quality metric, and discard ones that would collide with the bin, the table, or neighboring objects.
- Move to a pre-grasp pose just above or beside the object, fingers open, aligned with the chosen approach direction.
- Approach along that direction, close the jaws (or pull the vacuum), confirm contact, and lift — watching force or suction feedback to catch a slip early.