Geometric backtracking for combined task and path planning in robotic systems

Planners for real, possibly complex, robotic systems should not only reason about abstract actions, but also about aspects related to physical execution such as kinematics and geometry. We present an approach in which state-based forward-chaining task planning is tightly coupled with sampling-based motion planning and other forms of geometric reasoning. We focus on the problem of geometric backtracking which arises when a planner needs to reconsider geometric choices, like grasps and poses, that were made for previous actions, in order to satisfy geometric preconditions of the current action. Geometric backtracking is a necessary condition for completeness, but it may lead to a dramatic computational explosion due to the systematic exploration of the space of geometric states. In order to deal with that, we introduce heuristics based on the collisions between the robot and movable objects detected during geometric backtracking and on kinematic relations between actions. We also present a complementary approach based on propagating explicit constraints which are automatically generated from the symbolic actions to be evaluated and from the kinematic model of the robot. We empirically evaluate these dierent approaches. We demonstrate our planner on a real advanced robot, the DLR Justin robot, and on a simulated autonomous forklift. GeRTSAUN