Robotic Manipulation Leveraging Plane Contacts
July 1th :
Title： Robotic Manipulation Leveraging Plane Contacts
Lecturer：Jia Pan Assistant Professor
Time：13:30 – 16:00,July 1, 2015
Location：Room 201, Mathematics Building
General robotic manipulation is challenging. To successfully manipulate an object, the robot must first accurately estimate the object's pose from the sensor data, and then find a good grasp of the object via grasp planning, and finally use motion planning to compute a trajectory to approach the object. In some cases, the manipulation cannot be achieved in a single grasp due to cluttered environments or kinematics singularities, and regrasping is necessary for these cases. All these components of manipulation, including pose estimation, grasp planning, motion planning, and regrasp planning, can be difficult to solve.
In this talk, we will discuss about how to use plane contacts to tackle some of challenges in robotic manipulation.
The first part is about pose estimation. We present a new approach which effectively estimates the target object's pose given its partial observation, by leveraging two facts: 1) in most manipulation scenarios, the object is posed in a stable pose sitting on flat support surfaces, and 2) many household objects can only keep stable on a planar surface under a small set of poses. In this way, we can reduce the 6-dimensional search of pose estimation into a set of 3-dimensional search problem, and thus greatly improve the accuracy and efficiency of the estimated pose.
The second part is about regrasping. Similar to previous work, we use one intermediate location for the temporary placement between two pickups. However, unlike previous work which assumed this intermediate location to be horizontal, we show that a tilted surface can improve the object's stability and manipulability while being regrasped, and thus increase the flexibility of regrasping.
Jia Pan is a Research Assistant Professor in the Department of Computer Science, the University of Hong Kong. His research area is intelligent grasping and manipulation, including motion planning, deformable object manipulation, learning from demonstration, reinforcement learning and planning with uncertainty.