A Transition-Aware Method for the Simulation of Compliant Contact with Regularized Friction


Our method effectively simulates
contact with friction in manipulation tasks. We simulate contact using a continuous
model of compliant forces with regularized friction. This allows us to
cast the problem as a system of ODEs. We evaluate the performance of implicit
integrators for these stiff systems and designed a method that allows the
implicit Newton solves to succeed even when using large time steps. Implicit integrators use Newton’s method
to solve for the next V. With friction the residual’s slope changes drastically
between sliding and stiction. In this example, we start from V equals V0, in
the sliding region, and try to reach V* in stiction. At V0 Newton’s method
follows the slope and arrives at V1, which is on the other side of
the stiction zone. At V1, we follow the slope again and land on V2, which is on
the original side. This will only oscillate between V1 and V2, and never
converge. Let’s start again at V0, and make a slight modification this time. Our
method uses the system’s dynamics to backtrack the Newton step so that when
it tries to jump across, it stops in the stiction region. From here the slope is
more accurate and we find the route quickly. Consider a simple gripper grabbing and
shaking a mug. Strong disturbances during a manipulation task are emulated by
shaking the mug vertically. When the grasp transitions between sliding and stiction, a traditional implicit Euler scheme needs to take smaller time steps
to ensure the convergence of the implicit Newton solve. This causes the
short pauses in the video along with the jagged plateaus in the red line. Our transition aware line search can
take larger fixed time steps and thus runs more smoothly and faster. We evaluated the performance of our
TAMSI method on a complex manipulation tasks using a 16-dof
Allegro hand with multiple points of contact.
We demonstrate the robustness of our TAMSI method to simulate a manipulation
task with vigorous shaking motions emulating disturbances. A KUKA IIWA arm is
commanded to pick up a bottle, execute a shaking motion, and set it down at a
specified position, where force feedback is used to assess grasp quality as well
as task completion. We show that our framework is ideally suited for
prototyping controller processes for manipulation that seamlessly transfer to
reality.

Leave a Reply

Your email address will not be published. Required fields are marked *