Reliable robotic manipulation in everyday environments remains difficult because robots must interact with objects that differ in shape, weight, and frictional properties.
Manipulation tasks in such unstructured environments often require more than stable grasping, including reorientation of the object within the grasp.
A common example is pivoting, a reorientation task in which the object rotates within the fingers while translation is prevented.
This requires information about contact forces and proximity to slip at the fingertip, which can be provided by tactile sensing.
Recent work has shown that tactile data can be used to predict a frictional safety margin for slip caused by shear force.
However, in tasks involving rotation, such as pivoting, slip depends on both shear force and torque.
In this thesis, a generalized safety margin is introduced for pivoting by extending the safety margin estimation to also account for torque.
This is done using the Limit Surface, which describes the boundary between sticking and slipping under combined force and torque.
An experimental setup was built to measure slip under different force and torque combinations, from which the Limit Surface curves were fitted.
Using those curves and the displacement field from the tactile sensor, deep learning models were then trained to predict the contact forces and the generalized safety margin.
Validation on unseen data and experiments on grasp control and pivoting showed that tactile sensing can be used to predict and regulate friction during pivoting.
This thesis extends the frictional safety margin approaches of previous studies from translational loading to pivoting.
Estimating and regulating friction during pivoting could help robots perform a wider range of manipulation tasks in unstructured environments. ... Read more