Nonprehensile Manipulation

• Also called: non-prehensile manipulation, graspless manipulation
• Category: Planning & manipulation › Manipulation
• Tags: manipulation, pushing, contact, friction, mechanics

Moving or reorienting objects without grasping them, for example by pushing, sliding, tilting, rolling, pivoting or throwing, relying on friction, gravity and dynamics instead of a firm hold.

Definition

Nonprehensile manipulation is manipulation in which the robot changes the state of an object without holding it in a grasp. Examples include pushing a box across a table, sliding a thin book to the table edge so it can be picked up, tilting a heavy object onto its edge, rolling a cylinder, pivoting an object against a wall, batting or throwing. In a prehensile grasp the hand fully constrains the object, so it moves with the hand; in nonprehensile manipulation the object is only partly constrained, and its motion depends on friction, gravity, inertia and contact with the environment.

Because the object is not held, the robot has to predict how the object will respond to contact, which makes mechanics central. Matthew Mason's work on the mechanics of pushing in the 1980s established the foundations: for a slow push, the direction an object rotates depends on the center of friction and the friction cone at the pusher contact, and a pushing contact can be designed to be stable so that the object moves with the pusher without slipping. The limit surface describes the relationship between the friction wrench and the sliding motion of an object on a surface. These analytic models depend on friction coefficients and pressure distributions that are hard to know in advance, which is why many recent systems learn pushing models or residual corrections to analytic ones from data.

Nonprehensile manipulation overlaps with, but differs from, two neighboring terms. Quasi-static manipulation is an assumption about speed, that inertial forces can be ignored, and slow pushing is both quasi-static and nonprehensile. Dynamic manipulation relies on inertial forces, and throwing or batting are both dynamic and nonprehensile. Grasping itself can be quasi-static or dynamic, so these are separate axes. Nonprehensile actions are also used as a step before grasping, to move an object into a pose where a grasp becomes possible, which is closely related to extrinsic dexterity.

Why it matters

Many objects are too large, too flat, too heavy or too cluttered to grasp directly, and a simple gripper or even a bare paddle can do useful work by pushing and sliding. Nonprehensile skills widen what simple hardware can do, but they require good contact models or learned dynamics because the object is never fully under the robot's control.

Key applications

• Pushing objects to goal poses in cluttered workspaces
• Pre-grasp manipulation such as sliding objects to a table edge or away from walls
• Toppling, tilting and pivoting heavy or large items
• Sorting and singulation in bins and on conveyors
• Throwing, batting and other dynamic tasks
• Manipulation with simple end effectors such as paddles and fences

In practice

A logistics robot needs to pick a flat package lying flush against a bin wall where the gripper cannot reach under it. The planner first pushes the package away from the wall with the closed gripper, using a pushing model to choose a contact point and direction that translate the package without rotating it much. A camera tracks the package after each push, and the controller adjusts the next push if the package rotated more than predicted. Once a gap opens, the gripper slides one finger under the edge and grasps. The team logs the predicted and observed package poses to refine the friction parameters of the pushing model.

Common misconceptions

• Nonprehensile means dynamic — slow pushing is nonprehensile and quasi-static, while throwing is nonprehensile and dynamic
• Pushing is simple because it needs no grasp — the outcome depends on friction and pressure distribution, which are usually uncertain, so prediction is harder than for a grasped object
• Nonprehensile manipulation is only a workaround for poor grippers — it is often the most efficient way to handle large, flat or clustered objects even with capable hands

Who should know this

• Manipulation researchers and engineers
• Warehouse and logistics automation engineers
• Motion planning engineers working with contact
• Robot learning researchers training pushing and sliding policies

Related technologies

• Limit Surface Models
• Push Planning
• Contact-Implicit Trajectory Optimization
• Learned Dynamics Models
• Physics Simulators

Related terms

• Quasi-Static Manipulation
• Dynamic Manipulation
• Extrinsic dexterity
• Friction Cone
• Contact mode
• Kinodynamic Planning

Related projects

• LeRobot: HuggingFace ML library for real-world robotics in PyTorch
• Isaac Lab: NVIDIA robot learning framework: one task API across Newton, PhysX and Isaac Sim
• NVIDIA Isaac GR00T: NVIDIA open foundation model for humanoid robot reasoning and skills
• OpenVLA: 7B-parameter Vision-Language-Action model for robotic manipulation
• Drake: MIT robotics toolbox for model-based dynamics, optimization, and control system design
• OpenArm: 7DOF humanoid arm for physical AI research in contact-rich environments

Videos

• Humans, Robots, and Non-Prehensile Manipulation: Brian Douglas