Preprocessing-based Methods for Robotic Manipulation

Abstract

Robotic manipulation is a key problem for several applications such as welding, pick-and-place, and automated assembly. However, motion planning for manipulation can be computationally expensive as it requires planning in the high-dimensional configuration space of the manipulator. Additionally, task-specific constraints such as strict time limits or constraints on end-effector motion further increase the complexity of this problem. In order to solve the planning problem tractably, several methods heavily rely on preprocessing, in which relevant information about the planning problem is gathered beforehand in order to enhance the performance of the planner. In this thesis, we propose algorithmic and system-level contributions to preprocessing-based methods for robotic manipulation.

In the first part of the thesis, we demonstrate how planning for constrained manipulation can be sped up by adding additional actions to the action set of a search-based planner. Macro-actions are ordered combinations of primitive actions and can help make faster progress towards the goal and reduce planning time. However, this introduces a trade-off as additional actions will increase the branching factor of the search. We leverage preprocessing to compute a set of macro-actions that provide probably approximately correct (PAC) bounds on the performance of the planner. We demonstrate the benefits of our approach on a container-opening task with a PR2 robot arm.

In the second part of the thesis, we present a planning and perception framework for intercepting projectiles using robotic manipulators. We focus on the problem of intercepting a projectile moving towards a robot equipped with a manipulator holding a shield. To successfully perform this task, the robot needs to (i) detect the incoming projectile, (ii) compute a trajectory that can intercept the projectile, and (iii) execute the found trajectory. These three steps need to be executed as fast as possible (<= 1 second in our setting) in order to maximize the number of episodes where the projectile is successfully intercepted. To this end, we propose a planning framework that constructs a trajectory database, wherein the individual trajectories can be used in order to intercept the projectile. Additionally, we also present a perception pipeline that continually provides state estimates of a fast-moving projectile within a short time window. We evaluate our approach both in simulation and on the physical PR2 robot with RGB-D cameras.