Extrinsic Dexterity

Using gravity, a tabletop, a wall, or arm swinging to let a simple gripper carry out complex manipulation.

Extrinsic dexterity was proposed by Nikhil Chavan-Dafle, Alberto Rodriguez, Matthew Mason, and colleagues in an ICRA 2014 paper: instead of relying on the fingers' own dexterous motion, a robot uses resources external to the hand — gravity, contact with a tabletop or wall, or dynamic arm motion — to adjust an object's position in the hand or complete a manipulation. Traditional dexterous manipulation mainly relies on coordinated finger movement in a multi-fingered hand. The original paper designed 12 regrasping motions for a simple gripper and ran over 1,200 trials across 3 objects, showing that even a simple gripper can accomplish a good deal of in-hand manipulation this way. The significance is that high-degree-of-freedom dexterous hands are expensive and hard to control, while making good use of the environment can greatly extend what a simple gripper can do. Later work has used learning methods to automatically discover tricks like this — for example, Wenxuan Zhou and David Held (CoRL 2022) used reinforcement learning to let a gripper learn to push a flat object with no graspable edge against a wall, tip it upright, and then grasp it, reaching a 78% success rate when transferred from simulation to a real robot.

ExampleA book lying flat on a table gives the gripper nothing to grab; the robot first pushes it against a wall to tip it upright, then grips it from the side.

Related

Non-prehensile Manipulation · In-hand Manipulation · Contact-rich Manipulation · Dexterous Manipulation · Dynamic Manipulation · Gripper

Sources

Extrinsic Dexterity: In-Hand Manipulation with External Forces (Chavan-Dafle et al., ICRA 2014)
Learning to Grasp the Ungraspable with Emergent Extrinsic Dexterity (Zhou & Held, CoRL 2022)