Most robots today operate under a "quasistatic" assumption: the belief that the world only moves when the robot moves it, and stops the moment the robot stops. In this paradigm, inertia is a nuisance to be suppressed, and gravity is a constant to be countered. While this works for slow-moving factory arms, it fails for the fluid, high-speed tasks humans perform daily—like flipping a pancake, whipping a charging cable over a desk, or catching a falling glass.
A growing body of research is pushing robotics past this "slow and steady" wall, moving toward dynamic manipulation where momentum, impact, and centrifugal forces are not just managed, but exploited.
The Quasistatic Wall
In traditional Task and Motion Planning (TAMP), as surveyed by Caelan Reed Garrett and Tom Silver, physics is often simplified to ensure stability. If a robot is moving a block, it assumes the block’s velocity is zero the instant the gripper stops.
However, as Harry Zhang and Ken Goldberg point out in Robots of the Lost Arc, you cannot "pick and place" a cable into a narrow gap if that gap is out of reach. You have to flick it. This requires a transition from static geometry to temporal dynamics, where the robot must reason about the "arc" of the object’s motion after the robot has finished its own stroke.
Exploiting Momentum and Inertia
To master high-speed tasks, robots must account for forces that quasistatic models ignore: centrifugal, Coriolis, and inertial forces. Noémie Jaquier and Sylvain Calinon demonstrate that by tracking "manipulability ellipsoids" that include mass and inertia, robots can plan motions that are significantly more efficient for high-speed tasks like pushing or assembly.
This shift is most visible in "contact-rich" maneuvers. For example, Leidi Zhao and Bo Ouyang developed a framework for nunchaku flipping. This isn't just about moving an arm; it’s about managing the high-frequency state transitions of a multi-link object where the dynamics are dominated by momentum. Similarly, the Robot Air Hockey testbed introduced by Caleb Chuck forces agents to deal with objects that slide, bounce, and roll—environments where the "quasistatic" state is non-existent.
The Challenge of Deformable Objects
Linear Deformable Objects (DLOs) like ropes and power cables represent the "final boss" of dynamic manipulation. Because they have infinite degrees of freedom, you cannot model them as simple rigid bodies.
Mingrui Yu and Xiang Li have shown that while offline models can handle large deformations, they often fail during high-speed movement because they ignore the cable's own weight and inertia. By incorporating non-negligible dynamics, robots can precisely position heavy industrial cables by accounting for the "swing" and "bend" that occurs during rapid transport.
Learning Through Interaction
If the physics are too complex to model analytically, can we learn them?
Abdeslam Boularias and Changkyu Song use differentiable physics simulations to "learn" an object's mass and friction from just a few seconds of sliding. This allows the robot to perform impulsive actions—like a flick—that rely on the object's momentum to reach a target.
On the hardware side, platforms like the TriFinger, developed by Manuel Wüthrich and Felix Widmaier, provide the necessary "transparency" (backdrivability) to safely practice these high-speed interactions. This has led to breakthroughs like SWIFT, where Yunchao Yao and Jeffrey Ichnowski trained a soft robotic hand to perform pen spinning—a task that requires managing complex, contact-rich dynamics within millisecond timeframes.
Why It Works
The intuition behind this research direction is a move from position control to force and momentum control.
- Anticipation: Instead of reacting to where an object is, the robot plans for where the object's momentum will carry it (e.g., catching).
- Energy Transfer: The robot uses its own acceleration to "inject" energy into an object (e.g., throwing or whipping).
- Compliance: Using soft hardware or "residual force" metrics, as explored in the Residual Force Polytope work, allows the robot to absorb impacts and resist external disturbances that would shatter a rigid, quasistatic system.
By embracing the "messy" physics of the real world, researchers are moving toward robots that don't just move objects, but dance with them.