Dynamic Manipulation

Manipulation that deliberately exploits velocity, inertia, and gravity, such as throwing, flinging, catching, or tossing.

Dynamic manipulation means a robot deliberately exploits an object's velocity, inertia, gravity, centrifugal force, or other dynamic effects to complete a task, such as throwing, flinging, catching, or slapping. The counterpart is quasi-static manipulation, where motion is slow enough that inertia can be ignored and the system is approximately in static equilibrium at every moment — most pick-and-place falls into this category. One of the earliest systematic studies in robotics came from Kevin Lynch and Matthew Mason (IJRR 1999), who showed that even a simple arm with only one or two joints could control an object's state using rolling, sliding, and throwing. Dynamic manipulation is faster and can move objects beyond the arm's reach, but once an object leaves the hand it cannot be corrected, so accurate dynamics prediction is essential, and perception and control latency matter much more at high speed. Recent work has combined it with data-driven learning, such as TossingBot learning to throw objects into a bin, FlingBot learning to fling cloth open, and IRP learning to whip a rope to hit a target.

ExampleFlingBot grips two corners of a piece of cloth with both arms and flings it forward to spread out a tangled bundle — much faster than smoothing it out bit by bit, and able to unfold cloth larger than the arms' own reach.

Related

Non-prehensile Manipulation · Deformable Object Manipulation · Garment Manipulation · Extrinsic Dexterity · Quasi-Static Assumption · Manipulation

Sources

Dynamic Nonprehensile Manipulation: Controllability, Planning and Experiments (Lynch & Mason, IJRR 1999)
FlingBot: The Unreasonable Effectiveness of Dynamic Manipulation for Cloth Unfolding (CoRL 2021)
TossingBot: Learning to Throw Arbitrary Objects with Residual Physics