Past Events

PhD Defense

Scalable, Physics-driven Model Identification for Robust Robotic Manipulation


Download as iCal file

Thursday, March 25, 2021, 03:00pm - 05:00pm


Speaker: Changkyu Song

Location : Remote via Zoom


Professor Abdeslam Boularias (Advisor)

Professor Jingjin Yu

Professor Mridul Aanjaneya

Professor Tucker Hermans (external member)

Event Type: PhD Defense

Abstract: Manipulating all types of objects is still challenging in robotics. Successful manipulation planning requires accurate geometric and mechanical models of objects as a precondition. In warehouses and factories, manipulated objects are typically known in advance, with their CAD models obtained from full 3D scans along with other mechanical properties like mass and friction in the workspace. However, recent research efforts in grasping and manipulation rather focus on the tasks where object models are unavailable. For example, the robot cannot obtain full 3D scans and mechanical properties of unknown objects inside a drawer or in a pile of clutter. Furthermore, in such circumstance, nonprehensile manipulation of objects is preferred over the traditional pick-and-place approach because an object may not be easily grasped by the robot, due to the design of the end-effector, the size of the object, or the obstacles surrounding the manipulated object. Recently, combined pushing and grasping approaches have shown success where traditional grasp planners fail, and they work well under uncertainty. First, I propose the probabilistic methods to infer geometric and mechanical models of unknown objects by leveraging non-prehensile robot actions. The proposed methods utilize physics simulations of all the physical interactions between the objects and between the robot and the manipulated objects. Non-prehensile manipulation on unknown objects, such as pushing and sliding, and physics reasoning help infer the shape and mechanical properties of the objects by replaying hypothesized object models in simulation. Then, the identified probabilistic models enable probabilistic manipulation planning. I propose the probabilistic manipulation planning methods using the identified object models. The proposed planning methods shows successful pre-grasp sliding manipulation by selecting a stable goal configuration and finding the most robust action considering all possible hypothesized models. Finally, I propose the nonprehensile manipulation on multiple objects that allows the robot to rearrange the objects much efficiently. The nested nonprehensile manipulation actions reduce the length of the end-effector’s trajectories by manipulating multiple objects simultaneously. The nested rearrangement search algorithm proposes an efficient way to explore the object interactions in combinatorially large configuration space of multiple objects.


Join Zoom Meeting

Join by SIP

Meeting ID: 983 2033 3363
Password: 833835

One tap mobile
+16465588656,,98320333363# US (New York)
+13017158592,,98320333363# US (Washington DC)

Join By Phone
+1 646 558 8656 US (New York)
+1 301 715 8592 US (Washington DC)
+1 312 626 6799 US (Chicago)
+1 669 900 9128 US (San Jose)
+1 253 215 8782 US (Tacoma)
+1 346 248 7799 US (Houston)
Meeting ID: 983 2033 3363
Find your local number:

Join by Skype for Business