AI Planning meets Production Logistics Francesco Leofante Imperial - - PowerPoint PPT Presentation
AI Planning meets Production Logistics Francesco Leofante Imperial College London, United Kingdom FMAIL 2019 Bergen, Norway December 2, 2019 A (Nearly Epic) Tale of Courage, Passion, Amazing Victories and Major Defeats Francesco Leofante
AI Planning meets Production Logistics
Francesco Leofante
Imperial College London, United Kingdom FMAIL 2019 Bergen, Norway December 2, 2019 A (Nearly Epic) Tale of Courage, Passion, Amazing Victories and Major Defeats
Francesco Leofante
Imperial College London, United Kingdom FMAIL 2019 Bergen, Norway December 2, 2019 Starring...
Francesco Imperial College London Verification of Autonomous Systems Group @ Imperial
https://fraleo.github.io/
Starring...
Erika ´ Abrah´ am RWTH Aachen University Armando Tacchella University of Genoa Gerhard Lakemeyer RWTH Aachen University Tim Niemueller X - The Moonshot Factory
Agenda
AI Planning
The RoboCup Logistics League
Concluding remarks
Before we go into the details...a little teaser
Agenda
AI Planning
The RoboCup Logistics League
Concluding remarks
Why?
“A goal without a plan is just a wish”
from “50 Ways to Lose Ten Pounds” (1995) by Joan Horbiak, p. 95
What is planning anyway?
Planning is... model-based autonomous behavior
What is planning anyway?
Planning is... model-based autonomous behavior That is, given a model that represents
an initial situation I, the actions A that can be performed and a goal G
what are the actions that I should perform to reach G from I?
What is planning anyway? (cont.d)
Depending on the type of actions we are dealing with, we can define several forms of planning,
Path planning Motion planning Task planning ...
What is planning anyway? (cont.d)
Depending on the type of actions we are dealing with, we can define several forms of planning,
Path planning Motion planning Task planning ...
Here: task planning
Solving planning problems
Different techniques for different representations...
heuristic search in a graph symbolic search reductions to other problems...
Solving planning problems
Different techniques for different representations...
heuristic search in a graph symbolic search reductions to other problems...
Here: reductions to satisfiability checking
Planning as SAT
Let F and A be the sets of state and action variables. Let X = F ∪ A and X′ = {x′ : x ∈ X} be its next state copy. A planning problem is a triple of formulas Π = I, T, G where
I(F ) represents the set of initial states T(X, X′) describes how actions affect states G(F ) represents the set of goal states
Planning as SAT
Planning formula ϕ(Π,k) := I(F0) ∧
k−1
T(Xi, Xi+1) ∧ G(Fk) How to choose k?
start with k = 1 increase until ϕ(Π,k) SAT or upper bound is reached.
ϕ(Π,k) is sat iff there exists a plan with length k
in that case, a plan can be extracted from the satisfying assignment
Planning as SAT - limitations
Planning as SAT - limitations
hasFuel(robot): y or n?
Planning as SAT - limitations
Planning as SAT - limitations
minimize total time fuel(robot) > 5 hasMoney(robot)
Step 1: Planning Modulo Theories
How to overcome (some) limitations of SAT?
Satisfiability Modulo Theories
SMT: Boolean combinations of constraints expressed in decidable fragments of first-order theories
Step 1: Planning Modulo Theories
How to overcome (some) limitations of SAT?
Satisfiability Modulo Theories
SMT: Boolean combinations of constraints expressed in decidable fragments of first-order theories Integrated approach for new classes of planning problems:
numeric planning (costs, rewards...) temporal planning (durations, deadlines) planning with resources (raw materials)
Step 2: Optimal Planning Modulo Theories
Can we ask for more?
compute a plan that achieves a goal compute a plan that achieves a goal minimizing costs/time
Step 2: Optimal Planning Modulo Theories
Can we ask for more?
compute a plan that achieves a goal compute a plan that achieves a goal minimizing costs/time Optimization Modulo Theories
OMT: optimize objective function subject to SMT formulas
Agenda
AI Planning
The RoboCup Logistics League
Concluding remarks
The RoboCup Logistics League
BS RS 1 RS 2 RS 2 CS 2 Planning & Execution Competition for Logistics Robots in Simulation [Niemueller et al., 2015]
Planning & Execution Competition for Logistics Robots in Simulation [Niemueller et al., 2015]
Planning for the RCLL: what’s hard?1
Representation
state space languages and scalability?
Execution
changing environment temporary robot/machine failures changes in task definitions
Representational challenges
domain representation: over 250 product configurations possible! need a language that can model crucial aspects of the domain:
resources: how many token do I need to produce X? durations: how long does it take to produce X? deadlines: deliver X before... time and space dependencies
Solving the RCLL is not easy:
heuristic search does not scale rule-based approaches work, but no formal guarantees!
Can I panic now?
Exploration tasks
Exploration tasks
robot i
Exploration tasks
robot i location k location l
Exploration tasks
Step j
Exploration tasks
Step j posi,j+1 = l di,j+1
Exploration tasks
Step j posi,j+1 = l di,j+1 posi,j+2 = k di,j+2
Exploration tasks
Step j posi,j+1 = l di,j+1 posi,j+2 = k di,j+2
The move action for each robot i and step j can be encoded as
∀i.∀j.
Z
Z
lk
A (less) simple example: C0 production
1Retrieve base with cap from shelf at CS
2Prepare CS to retrieve cap
3Feed base into CS
4Discard cap-less base
5Prepare BS to provide black base
6Retrieve base from BS
7Prepare CS to mount cap
8Feed black base to CS
9Retrieve black base with cap from CS
10 Prepare DS for slide specified in order 11 Deliver to DS BS CS 2 C S BS DS Standard encodings
s0 s1 s2 s3 s4
a1
→
a2
→
a3
→
a4
→
Standard encodings
s0 s1 s2 s3 s4
a1
→
a2
→
a3
→
a4
→
Standard encodings
s0 s1 s2 s3
a1∧a2
→
a3
→
a4
→
Standard encodings
s0 s1 s2 s3
a1∧a2
→
a3
→
a4
→
Standard encodings
s0 s1 s2 s3
a1∧a2
→
a3
→
a4
→
Our solution: long-distance dependencies
[Leofante et al., 2018a] Our solution: long-distance dependencies
[Leofante et al., 2018a] Our solution: long-distance dependencies
[Leofante et al., 2018a] Our solution: long-distance dependencies
[Leofante et al., 2018a] Our solution: long-distance dependencies
[Leofante et al., 2018a]s0 s1 s2 s3 s4
Our solution: long-distance dependencies
[Leofante et al., 2018a]s0 s1 s2 s3 s4
Our solution: long-distance dependencies
[Leofante et al., 2018a]s0 s1 s2 s3 s4
Our solution: long-distance dependencies
[Leofante et al., 2018a]s0 s1 s2 s3 s4
. . .
Integrated planning and execution
[Leofante et al., 2018a]Model
Agenda
AI Planning
The RoboCup Logistics League
Concluding remarks
Domain-independent OMT planning
We implemented a domain-specific planner for the RCLL
Domain-independent OMT planning
We implemented a domain-specific planner for the RCLL
Cool approach but...Does it scale on other planning problems?
(The Guardian)
Optimal Planning Modulo Theories
[Leofante et al., 2019, Leofante, 2019]Major limitations of [Leofante et al., 2018a]:
tailored for the RCLL
plans have specific structures, horizon is known, in general blow-up heavy use of domain knowledge
how to infer type-abstractions automatically?
Optimal Panning Modulo Theories
[Leofante et al., 2019, Leofante, 2019]I T GR TR TC n n + 1 abstraction to boost OMT planning Implemented in OMTPlan, available at https://github.com/fraleo/OMTPlan
Summing up
Achieved so far implemented a domain-specific OMT planner for the RCLL integrated OMT planning with online execution and monitoring developed new encodings inspired by challenges faced in the process
Summing up
Achieved so far implemented a domain-specific OMT planner for the RCLL integrated OMT planning with online execution and monitoring developed new encodings inspired by challenges faced in the process
Are we done?
Summing up
Achieved so far implemented a domain-specific OMT planner for the RCLL integrated OMT planning with online execution and monitoring developed new encodings inspired by challenges faced in the process
Are we done? There’s still a lot to do to ease applicability of (OMT) planning in logistics!
Open challenges
Some open problems kick-start discussions:
scalability: dedicated solvers? uncertainty: expressive models or ”online” solutions? integrated task and motion planning: hybrid models? explainability: why did you do that? ...many more!
Some references I
Leofante, F. (2019). The OMTPlan planner: system description. (in preparation). Leofante, F., ´ Abrah´ am, E., Niemueller, T., Lakemeyer, G., and Tacchella, A. (2018a). Integrated synthesis and execution of optimal plans for multi-robot systems in logistics. Information Systems Frontiers. Leofante, F., ´ Abrah´ am, E., and Tacchella, A. (2018b). Task planning with OMT: an application to production logistics. In Proc. of IFM, pages 316–325. Leofante, F., Giunchiglia, E., ´ Abrah´ am, E., and Tacchella, A. (2019). Optimal planning modulo theories. (under review). Niemueller, T., Lakemeyer, G., and Ferrein, A. (2015). The RoboCup Logistics League as a benchmark for planning in robotics. In Proc. of PlanRob@ICAPS’15.