Simultaneous Tactile Estimation and Control
of Extrinsic Contact
Sangwoon Kim1, Devesh K. Jha2, Diego Romeres2, Parag Patre3, and Alberto Rodriguez1
1MIT, 2Mitsubishi Electric Research Laboratories, 3Magna International Inc.
Overview
We explore how to use tactile sensing to manipulate the grasped object in contact with an environment. We propose a factor graph-based 'Simultaneous Tactile Estimator-Controller' to jointly estimate and control the contact state.
Abstract
We propose a method that simultaneously estimates and controls extrinsic contact with tactile feedback. The method enables challenging manipulation tasks that require finesse at the contact, such as stably placing an unknown object on a surface or even on a thin rod standing upright. A factor graph-based estimator-controller framework fuses a sequence of tactile measurements to estimate the contact state, such as the location of extrinsic contact between the grasped object and the environment and the intrinsic wrench exerted from the tactile sensor to the object. The same architecture also plans the robot motion to allow the collection of measurements from different poses while satisfying control objectives such as minimizing the amount of intrinsic wrench or minimizing tangential force at the extrinsic contact. The proposed method shows a contact localization error of less than 1 mm. It also reduces the tangential force at the extrinsic contact, which allows localizing the contact without slipping, even in a slippery environment. The proposed method also works for multiple contact formations (point, line, patch contact) and detects transitions between them.
Contact: Sangwoon Kim (sangwoon@mit.edu)
Method
We use a factor graph to simultaneously estimate and control the contact state with tactile feedback. Both estimation and control are computed within a single integrated architecture.
Estimation (left): Takes in the measurements and constraints to estimate the past and current contact states. The measurements include the tactile sensing from the GelSlim fingers and the gripper pose from the robot's forward kinematics. The constraints include the geometric constraint that the contact location should stay fixed over time and the physical constraint that there should be no torque exerted at the contact location.
Control (right): Takes in the control objectives and desired rotation to compute the robot's motion plan. The control objective that we used in our case is to minimize the elastic potential energy due to the finger deformation.
Example Results
Placing an Object
On Thin Rods
