Hierarchical Task and Motion Planning in Now Leslie P. Kaelbling and Tomas Lozano-Perez Department of Computer Science & Engineering, MIT Kai Liu June 19, 2013

1 Outline

Outline  Introduction: Motivation & Goal  Representation: Goals & Operations  Main ideas  Algorithms  Experiment  Conclusions

2 Introduction

Introduction Characteristics of problem to solve: Long time horizons Many continuous dimensions E.g. Continuous geometry Discrete aspects No terrible outcome: reversible Geometry is not complicated: not tight => strict optimality is not crucial

Introduction Goal: outline an approach to design a system that can work effectively with non-determinism in the environment, with two key properties: Aggressively hierarchical Operates in continuous geometry without any discretization of the state or action spaces

3 Representation

Representation  Fluents: Describe the logical aspects of domain In(O,R) : has value True if object O is entirely contained in region R, otherwise False Overlaps(O,R) : has value True if object O overlaps region R, otherwise False ClearX (R, Os) :has value True if region R is not overlapped by any object except those in the list Os, otherwise False Holding() : has value None if the robot is not grasping an object; otherwise the object being grasped Clean(O) : has value True if object O is clean and otherwise False

Representation  World States: a completely detailed description of both the geometric and non- geometric aspects of a situation.  Goals: A goal for our planning and execution system is a set of world states, described using a conjunction of ground fluents. The goal of having object a to be clean and in the washer can be articulated as: In(a, washer) = True ^ Clean(a) = True ;

Representation Operators: One operation is characterized by one or more operator descriptions. Define operators in a STRIPS-style form (finite domain): F(A1,….,An) = V : exists: B1,….,Bk pre: Φ1,…., Φm sideEffects: Ψ1,….., Ψ l prim: Π cost: c Abstract operator To operate in infinite domain, Suggester: to restrict domains for existential variables.

4 Main ideas

Main ideas  Hierarchical structure Expand Goal to several subgoals Speed construction of planning Subgoal is an abstract operator  Planning in Now Maintain left expansion of plan tree execute primitives plan as necessary  Why ? To handle Non-determinism !!

Main ideas

4 Algoritms

Algorithms

4 Experiment

Experiment Apply HPN method to several different configurations of objects in the simple domain right Also construct a larger ‘household’ domain with 6 connected room, with random ‘junk’ objects

4 Conclusions

Conclusions  Outline a general strategy for hierarchical planning and execution for task and motion planning.  Very suited to re-planning approaches to stochastic domains Based on the outcome of executing a primitive, any level above this step can be replanned  The potential for speedup relies on good domain-dependent choices  Improvement Apply learning algorithms to improve these choices

T hank you

Q &A