Manipulation Among Movable Objects for Pick-and-Place Tasks in 3D Workspaces
Keywords: robot manipulation, motion planning, physics-based simulation, contact-rich manipulation
Abstract: In cluttered real-world workspaces, simple pick-and-place tasks for robot manipulators can be quite challenging to solve. Often there is no collision-free trajectory that allows the robot to grasp and extract the desired object from the scene. This requires motion planning algorithms to reason about rearranging some of the movable clutter in the scene so as to make the task feasible. Our work focuses on solving these pick-and-place tasks in 3D workspaces where objects may tilt, lean on each other, topple, and slide. We formulate the problem as a search over a discrete graph -- vertices are configurations of all movable objects in the scene, and edges encode rearrangement actions taken to reach one configuration from another. The search solves a multi-agent pathfinding abstraction of the problem to generate candidate nonprehensile and prehensile rearrangement actions. In order to account for complex multi-body interactions in the scene, and to ensure that object-centric "interaction constraints" are satisfied during all rearrangement actions, the search queries a rigid-body physics simulator when evaluating actions. This helps us guarantee that the solution we find (i) does not make contact with immovable obstacles, and (ii) does not tilt movable objects beyond allowed limits nor makes them fall out of the workspace or move with high velocities. In addition, we introduce an easy-to-implement parallelisation scheme to deal with uncertainty in relevant object parameters (mass and coefficient of friction). We present an extensive evaluation of this approach in simulation on scenes of varying difficulty with a PR2 robot.
Submission Number: 13
Loading