Grasping and Manipulation

You pick up a full coffee mug without thinking twice. You do not drop it, you do not crush it, you do not squeeze so hard that coffee erupts from the top. Your hand and brain solved — in milliseconds and without conscious thought — a problem that has stumped robotics engineers for decades: reliable, general-purpose grasping. A robot tasked with picking up arbitrary objects in uncontrolled environments still fails routinely on tasks a five-year-old handles easily. This lesson explores why grasping is hard and how robotics engineers are tackling it.

What Makes a Grasp Stable?

A grasp is stable if the object cannot slip, rotate, or fall out of the hand under the forces it will experience during the task. Researchers formalize this with the idea of force closure: a grasp has force closure if the contact forces between fingers and object can resist any external force or torque applied to the object, in any direction. Force closure requires at least two contact points for a 2D object and typically three or more in three dimensions, arranged so that together they box the object in from multiple directions. A parallel-jaw gripper — two flat plates pressing from opposite sides — achieves force closure on a rectangular box. It fails on a sphere that can roll out sideways. The position of contact matters as much as the number of contacts. Contact at the widest part of an object, aligned with the center of mass, is usually more stable than contact near an edge.

Types of Grippers

The simplest and most common industrial gripper is the parallel-jaw gripper: two flat jaws driven by a single actuator that close symmetrically toward the center. It is fast, reliable, and easy to control, and it works well on boxes, cylinders, and any object with two flat parallel faces. The entire Amazon warehouse automation ecosystem is built largely around parallel-jaw grippers. A three-finger gripper adds a third jaw, allowing it to center-grasp cylindrical objects from any approach angle and handle a wider range of shapes. Multi-finger dexterous hands — robot hands with four or five fingers, each with multiple joints — can achieve the rich manipulation capabilities of a human hand: rolling a pen between fingers, turning a door knob, tying a knot. They are also expensive, fragile, and enormously difficult to control. Soft grippers made from silicone or rubber use air pressure to inflate finger-like chambers that wrap around and conform to irregular shapes. They exert gentle, distributed force — ideal for fragile objects like tomatoes, eggs, and electronics. A soft gripper does not know the exact shape of what it is picking up; it just squeezes gently and lets the object's own shape guide the grip.

Sensing for Grasping

A gripper without sensing is blind. It does not know whether it has gripped the object or whether the object is slipping. Modern robotic grippers increasingly include tactile sensors — arrays of pressure-sensitive elements on the finger surfaces, similar in concept to your skin's touch receptors — that tell the controller the force distribution across the contact area. Slip detection is critical: if the object begins to slide, the controller can increase grip force before a drop occurs. Without tactile sensing, the robot must either grip so hard that fragile objects break, or grip so lightly that heavy objects fall. Tactile feedback allows the grip force to be tuned to exactly what the object needs. Vision from a camera above or on the gripper tells the robot where the object is, its approximate size and orientation, and which approach angle is most likely to achieve a stable grasp. Combining camera-based grasp planning with tactile feedback during execution is the current state of the art for general grasping.

Why General Grasping Is Still Hard

Structured industrial settings sidestep most of the difficulty: every part arrives at the same position, same orientation, same size, on the same conveyor belt. The robot does not need to perceive or generalize — it executes the same programmed motion every time. Unstructured environments — a kitchen counter, a recycling bin, a hospital tray — present an enormous variety of objects in random positions and orientations, some occluded by others, some unknown to the robot. Grasping here requires perceiving an unknown object, generating a grasp candidate that is likely to be stable without exact geometric knowledge, executing the grasp, and detecting and recovering from failures. Deep learning has recently made significant progress on grasp planning from camera images: robots trained on millions of simulated or real grasp attempts have learned to predict whether a proposed grasp will succeed. But robustness in truly open-ended environments remains an active area of research.