Non-prehensile Manipulation

Moving an object by pushing, poking, flipping, or throwing it, instead of gripping it firmly.

Non-prehensile manipulation covers ways a robot changes an object's position or orientation without fully grasping it in a gripper or hand: pushing, sliding, poking, flipping, tossing, or propping it up. It is useful because some objects cannot be grasped at all — they're too large, too flat, flush against a wall, or blocked — and sometimes a push is simply faster than picking an object up and setting it down again. The difficulty is that how an object moves depends on friction and contact, which are hard to predict and control precisely. Mason, Lynch, and others did early work on the mechanics of pushing, and Ruggiero and colleagues published a survey of dynamic non-prehensile manipulation in RA-L in 2018. The Push-T benchmark, which asks a robot to push a T-shaped block to a target position, is a common testbed. The topic often comes up alongside extrinsic dexterity, where a robot uses the environment, such as a tabletop or wall, to help manipulate an object.

ExampleA flat card lying on a table can't be pinched up directly, so a robot first pushes it to the table's edge until part of it overhangs, then grips it from the side. Zhou and Held (CoRL 2022) used reinforcement learning to make a simple gripper push objects against a wall to flip them upright before grasping, reaching a 78% success rate on a real robot.

Also called

Nonprehensile Manipulation

Related

Extrinsic Dexterity · Contact-rich Manipulation · Dynamic Manipulation · Grasping · Push-T · Quasi-Static Assumption

Sources

Learning to Grasp the Ungraspable with Emergent Extrinsic Dexterity (Zhou & Held, CoRL 2022)
Nonprehensile Dynamic Manipulation: A Survey (Ruggiero et al., RA-L 2018)