kartik csig talk
TRANSCRIPT
Human-in-the-Loop Planning and
Decision Support
Subbarao KambhampatiArizona State University
1
CSIG Speaker Series
Adapted from a AAAI 2015 Tutorialrakaposhi.eas.asu.edu/hilp-tutorial/index.htm
Kartik TalamadupulaIBM T.J. Watson Research Center
AI’s Curious Ambivalence to Humans..
You want to help humanity, it is the people that you just can’t stand…
• Our systems seem happiest ..
• …either far away from humans
• …or in an adversarial stance with humans
What happened to Co-existence?
• Whither McCarthy’s advice taker?
• ..or Janet Kolodner’s house wife?
• …or even Dave’s HAL? • (with hopefully a less sinister voice) THIS TALK
HAAI in Planning
AI Planning: The Canonical View
4
A fully specifiedproblem--Initial state--Goals
(each non-negotiable)--Complete Action Model
The Plan
Human-in-the-Loop Planning
• In many scenarios, humans are part of the planning loop, because the planner:
• Needs to plan to avoid them• Human-Aware Planning
• Needs to provide decision support to humans• Because “planning” in some scenarios is too important to be
left to automated planners
• “Mixed-initiative Planning”; “Human-Centered Planning”; “Crowd-Sourced Planning”
• Needs help from humans• Mixed-initiative planning; “Symbiotic autonomy”
• Needs to team with them• Human-robot teaming; Collaborative planning
5
A Brief History of HILP• Beginnings of significant interest in the 90’s
• Under the aegis of ARPA Planning Initiative (and NASA)
• Several of the critical challenges were recognized
• Trains project [Ferguson/Allen; Rochester]
• Overview of challenges [Burstein/McDermott + ARPI Cohort]
• MAPGEN work at NASA
• At least some of the interest in HILP then was motivated by the need to use humans as a “crutch” to help the planner
• Planners were very inefficient back then; and humans had to “enter the land of planners” and help their search..
• In the last ~15 years, much of the mainstream planning research has been geared towards improving the speed of of plan generation
• Mostly using reachability and other heuristics
• Renaissance of interest in HILP thanks to the realization that HILP is critical in many domains even with “fast” planners
6
Dimensions of HIL Planning
7
Cooperation
Modality
Communication
Modality
What is
CommunicatedKnowledge Level
CrowdsourcingInteraction
(Advice from
planner to humans)
Custom Interface Critiques, subgoalsIncomplete Preferences
Incomplete Dynamics
Human-Robot
TeamingTeaming/Collaborat
ion
Natural Language
Speech
Goals, Tasks, Model
information
Incomplete Preferences
Incomplete Dynamics
(Open World)
“Grandpa Hates
Robots”
Awareness (pre-
specified
constraints)
Prespecified
(Safety / Interaction
Constraints)
No explicit
communication
Incomplete Preferences
Complete Dynamics
MAPGEN
Interaction
(Planner takes
binding advice from
human)
Direct Modification of
Plans
Direct modifications,
decision alternatives
Incomplete Preferences
Complete Dynamics
How do we adapt/adopt
modern planning technology for HILP?
8
PlanningThe Canonical View
9
Plan (Handed off
for Execution)
Full
Problem
Specification
PLANNER
Fully Specified
Action Model
Fully Specified
Goals
Completely Known
(Initial) World StateAssumption: Complete Action DescriptionsFully Specified PreferencesAll objects in the world known up frontOne-shot planning
Allows planning to be a pure inference problem
But humans in the loop can ruin a really a perfect day
Violated Assumptions:Complete Action Descriptions (Split knowledge)Fully Specified Preferences (uncertain users)Packaged planning problem (Plan Recognition)One-shot planning (continual revision)
Planning is no longer a pure inference problem
0000000000000000
000
0
0000000000000000
000
0
0000000000000000
000
0
0000000000000
Human-in-the-Loop Planning
10
Coordinate with Humans[IROS14, PlanRob15]
Replan for the Robot
[AAAI10, DMAP13]
Communicate with
Human(s) in the Loop
Open World Goals
[IROS09, AAAI10, TIST10]
Action Model Information
[HRI12]
Handle Human Instructions
[ACS13, IROS14]
Assimilate Sensor
Information
Full
Problem
Specification
PLANNER
Fully Specified
Action Model
Fully Specified
Goals
Completely Known
(Initial) World State
Sapa Replan
Problem Updates
[TIST10]
Human-in-the-Loop
Planning and
Decision Support
Goal
Manager
Challenges for the Planner1. Interpret what humans are doing
• Plan/goal/intent recognition
2. Decision Support• Continual planning/Replanning
• Commitment sensitive to ensure coherent interaction• Handle constraints on plan
• Plan with incompleteness• Incomplete Preferences• Incomplete domain models
• Robust planning with “lite” models• (Learn to improve domain models)
3. Communication• Explanations/Excuses
• Excuse generation can be modeled as the (conjugate of) planning problem
• Asking for help/elaboration• Reason about the information value
Challenge: Interpretation
•Recognize and interpret what the human needs
(goal, intent) and is currently doing (plan)
•Structure of “plans”:
• Assume Structure: Plan/Goal Recognition• Exploit/assume structured representation (plan)
• Easier to match planner’s expectation of structured input
• Restricts flexibility of humans; less knowledge specified
• Extract/Infer Structure: Infer Human Plans• Allow humans to use natural language
• Semi-structured and unstructured text
• Extract information from human-generated input
• Validate against (partial) model
• Iteratively refine recognized goals and plan
12
Plan/Goal Recognition for
Human-Robot Teaming
Talamadupula et al. – Arizona State University & Tufts University
Coordination in Human-Robot Teams Using Mental Modeling & Plan Recognition
BELIEF IN GOAL
Planning Conversation Robot Task Plans
Go here
Inferring Human (Team) Task Plans
• Integrate robots seamlessly in time-critical domains
• Lessen burden of programming and deploying robots
• Leverage the use of web-based planning tool (NICS)
14
Inferring Robot Task Plans from Human Team Meetings
Been Kim, Caleb Chacha and Julie Shah – MIT
Challenge: Decision Support
• Need to generate (and maintain) plans in the presence of human quirks and real-world considerations
– Partially specified preferences• Diverse plans
– Fully specified constraints / preferences• “Grandpa hates robots”
– Partially specified model of the world• Robust plans
– Changing world state and goals• Replanning
15
Problem Statement:
Given
the objectives Oi,
the vector w for convex combination of Oi
the distribution h(w) of w,
Return a set of k plans with the minimum ICP value.
Solution Approaches:
Sampling: Sample kvalues of w, and approximate the optimal plan for each value.
ICP-Sequential: Drive the search to find plans that will improve ICP
Hybrid: Start with Sampling, and then improve the seed set with ICP-Sequential
[Baseline]: Find kdiverse plans using the distance measures from [IJCAI 2007] paper; LPG-Speed.
1616
Generating diverse plans to handle unknown and
partially known user preferencesNguyen, Do, Gerevini, Serina, Srivastava & Kambhampati, 2012
Handling Partially Specified Preferences
17
Grandpa Hates RobotsKöckemann, Karlsson & Pecora
Fully Specified Constraints
• Compilation approach: Compile into a (Probabilistic) Conformant Planningproblem
• One “unobservable” variable per each possible effect/precondition
• Significant initial state uncertainty
• Can adapt a probabilistic conformant planner such as POND [JAIR, 2006; AIJ 2008]
• Direct approach: Bias a planner’s search towards more robust plans
• Heuristically assess the robustness of partial plans
• Need to use the (approximate) robustness assessment procedures
[Bryce et. Al., ICAPS 2012; Nguyen et al; NIPS 2013; Nguyen & Kambhampati, ICAPS 2014
Handling Partial Models of the World
19
Replanning for Changes in the World
A Theory of Intra-Agent ReplanningTalamadupula, Smith, Cushing & Kambhampati
20
Challenge: Communication
• Planner has to communicate with humans to:• Provide explanations
• Ask for help
• Providing Explanations• Excuses for failures
• Explanations for state facts
• Asking for Help• Symbiotic autonomy – humans doing what robots
cannot
• Generating detailed utterances to elicit effective help
21
Coming up With Good ExcusesGöbelbecker, Keller, Eyerich, Brenner & Nebel (2010)
Generating Excuses
22
Preferred ExplanationsSohrabi, Baier & McIlraith (2011)
Carnegie Mellon University Stephanie Rosenthal and Manuela Veloso 23
[Rosenthal, Veloso ,Dey: Ro-Man 2009, IUI 2010, JSORO 2012]
People are great sources
of information for robots‘Humans as Sensors and Effectors’
Symbiotic AutonomyRosenthal, Veloso et al.
Asking for Help – Symbiotic Autonomy
“Can you hold the elevator door open for me?”
“Can you point to where on this map?”
“Can you tell me when we reach the 4th floor?”
“Can you transfer this item from my basket?”
24
Asking for Help Using Inverse SemanticsTellex, Knepper, Li, Rus & Roy
Generating Utterances to Elicit Help
Case StudyHuman-Robot Teaming + Crowdsourcing
PLANNING FOR
HUMAN-ROBOT TEAMING
PLANNING FOR
CROWDSOURCING
INTERPRETATION
Open World Goals
Continual (Re)Planning
Plan & Intent Recognition
Activity Suggestions
Activity Critiques
Ordering Constraints
DECISION SUPPORTContinual (Re)Planning
Plan & Intent Recognition
Sub-goal Generation
Constraint Violations
Continual Improvement
COMMUNICATION Model Updates Sub-goal Generation
25
Dimensions of HIL Planning
26
Cooperation
Modality
Communication
Modality
What is
CommunicatedKnowledge Level
CrowdsourcingInteraction
(Advice from
planner to humans)
Custom Interface Critiques, subgoalsIncomplete Preferences
Incomplete Dynamics
Human-Robot
TeamingTeaming/Collaborat
ion
Natural Language
Speech
Goals, Tasks, Model
information
Incomplete Preferences
Incomplete Dynamics
(Open World)
“Grandpa
Hates Robots”
Awareness (pre-
specified
constraints)
Prespecified
(Safety / Interaction
Constraints)
No explicit
communication
Incomplete Preferences
Complete Dynamics
MAPGEN
Interaction
(Planner takes
binding advice from
human)
Direct Modification of
Plans
Direct modifications,
decision alternatives
Incomplete Preferences
Complete Dynamics
Recap
27
• HILP raises several open challenges for planning systems, depending on the modality of interaction between human and planner
• This talk:
– Discussed dimensions of variation of HILP tasks
– Identified the planning challenges posed by HILP scenarios
• Interpretation, Decision Support and Communication
– Outlined a few current approaches for addressing these challenges
rakaposhi.eas.asu.edu/hilp-tutorial