CSE 591: Human-aware Roboticsyzhan442/teaching/CSE591-HAR/...Human&aware*Robo.cs* 1 CSE 591:...
Transcript of CSE 591: Human-aware Roboticsyzhan442/teaching/CSE591-HAR/...Human&aware*Robo.cs* 1 CSE 591:...
Human-‐aware Robo.cs
1
CSE 591: Human-aware Robotics
Instructor: Dr. Yu (“Tony”) Zhang
Location & Times: CAVC 359, Tue/Thu, 9:00--10:15 AM Office Hours: BYENG 558, Tue/Thu, 10:30--11:30AM
Sep 22, 2016
This set of slides borrow from various online sources; it is used for educational purposes only.
Slides adapted from Henry Kautz and Oregon State University
Human-‐aware Robo.cs Approaches to goal/plan recogni>on • Consistency-‐based
– Hypothesize & revise – Closed-‐world reasoning – Version spaces
• Probabilis>c – Stochas>c grammars – Pending sets – Layered hidden Markov models – Policy recogni>on – Hierarchical hidden semi-‐Markov models – Dynamic probabilis>c rela>onal models – Dynamic Bayes networks – Example applica>on: Assisted Cogni>on
Can be complementary.. First pick the consistent plans, and check which of them is most likely (tricky if the agent can make errors)
3
Human-‐aware Robo.cs Hypothesize & Revise
• The Plan Recogni.on Problem C. Schmidt, 1978
Based on psychological theories of human narra>ve understanding
Men>on of objects suggest hypothesis
Pursue single hypothesis un>l matching fails
4
Human-‐aware Robo.cs Closed-‐world reasoning
• A Formal Theory of Plan Recogni>on and its Implementa>on Henry Kautz, 1991
• Infers the minimum set(s) of independent plans that entail the observa>ons
• Observa>ons may be incomplete
• Infallible agent • Complete plan library
5
Human-‐aware Robo.cs Version Space Algebra
• A sound and fast goal recognizer Lesh & Etzioni • Programming by Demonstra>on Using Version Space
Algebra Lau, Wolfman, Domingos, Weld.
• Recognizes novel plans • Complete observa>ons
• Sensi>ve to noise • Start with all goals
• Remove goals when ac>ons are not consistent
6
Human-‐aware Robo.cs Approaches to goal/plan recogni>on • Consistency-‐based
– Hypothesize & revise – Closed-‐world reasoning – Version spaces
• Probabilis>c – Stochas>c grammars – Pending sets – Layered hidden Markov models – Policy recogni>on – Hierarchical hidden semi-‐Markov models – Dynamic probabilis>c rela>onal models – Dynamic Bayes networks – Example applica>on: Assisted Cogni>on
Can be complementary.. First pick the consistent plans, and check which of them is most likely (tricky if the agent can make errors)
7
Human-‐aware Robo.cs Stochas>c grammars
– Huber, Durfee, & Wellman, "The Automated Mapping of Plans for Plan Recogni>on", 1994
– Darnell Moore and Irfan Essa, "Recognizing Mul>tasked Ac>vi>es from Video using Stochas>c Context-‐Free Grammar", AAAI-‐02, 2002.
CF grammar w/ probabilis>c rules
Successful for highly structured tasks (e.g. playing cards)
Problems: errors, context
Human-‐aware Robo.cs Pending sets
• A new model of plan recogni>on. Goldman, Geib, and Miller • Probabilis>c plan recogni>on for hos>le agents. Geib, Goldman
Explicitly models the agent’s “plan agenda” using Poole’s “probabilis>c Horn abduc>on” rules
Handles mul>ple concurrent interleaved plans & nega>ve evidence
Number of different possible pending sets can grow exponen>ally
Context problema>c? Metric >me?
Happen(X,T+1) ß Pending(P,T), X in P, Pick(X,P,T+1).
Pending(P’,T+1)ß Pending(P,T), Leaves(L), Progress(L, P, P’, T+1).
Human-‐aware Robo.cs Layered hidden Markov models
• N. Oliver, E. Horvitz, and A. Garg. Layered Representa>ons for Recognizing Office Ac>vity, Proceedings of the Fourth IEEE Interna.onal Conference on Mul.modal Interac.on (ICMI 2002)
Cascade of HMM’s, opera>ng at different temporal granulari>es
Inferen>al output at layer K is “evidence” for layer K+1
Human-‐aware Robo.cs Policy Recogni>on
• Tracking and Surveillance in Wide-‐Area Spa>al Environments Using the Hidden Markov Model. Hung H. Bui, Svetha Venkatesh and West.
• Bui, H. H., Venkatesh, S., and West, G. (2000) On the recogni>on of abstract Markov policies. Seventeenth Na>onal Conference on Ar>ficial Intelligence (AAAI-‐2000), Aus>n, Texas
Model agent using hierarchy of abstract policies (e.g. abstract by spa>al decomposi>on)
Compute the condi>onal probability of top-‐level policy given observa>ons
Compiled into DBN
Human-‐aware Robo.cs Hierarchical hidden semi-‐Markov models
Combine hierarchy (func>on call seman>cs) with metric >me
Compile to DBN Time nodes represent a
distribu>on over the >me of the next state “switch”
“Linear >me” smoothing – Research issues –
parametric >me nodes, varying granularity
l Hidden semi-‐Markov models (segment models) Kevin Murphy. November 2002.
l HSSM: Theory into Prac>ce, Deibel & Kautz, forthcoming.
Human-‐aware Robo.cs Dynamic probabilis>c rela>onal
models
• Friedman, N., L. Getoor, D. Koller, A. Pfeffer. Learning Probabilis>c Rela>onal Models. IJCAI-‐99, Stockholm, Sweden (July 1999).
• Rela>onal Markov Models and their Applica>on to Adap>ve Web Naviga>on, Anderson, Domingos, Weld 2002.
• Dynamic probabilis>c rela>onal models, Anderson, Domingos, Weld, forthcoming.
PRM -‐ reasons about classes of objects and rela>ons
Lajce of classes can capture plan abstrac>on
DPRM – efficient approximate inference by Rao-‐Blackwellized par>cle filtering
Open: approximate smoothing?
Human-‐aware Robo.cs
Dynamic Bayes Networks are “templates” for specifying the relation between the values of a random variable across time-slices !e.g. How is Rain at time t related to Rain at time t+1? We call them templates because they need to be expanded (unfolded) to the required number of time steps to reason about the connection between variables at different time points
Human-‐aware Robo.cs Special Cases of DBNs are well known in the literature
• Restrict number of variables per state – Markov Chain: DBN with one variable that is fully observable
– Hidden Markov Model: DBN with only one state variable that is hidden and can be es>mated through evidence variable(s)
• Restrict the type of CPD – Kalman Filters: DBN where the system transi>on func>on as well as the observa>on variable are linear gaussian
• The advantage of Gaussians is that the posterior distribu>on remains Gaussian
Human-‐aware Robo.cs Dynamic Bayes nets
• E. Horvitz, J. Breese, D. Heckerman, D. Hovel, and K. Rommelse. The Lumiere Project: Bayesian User Modeling for Inferring the Goals and Needs of Solware Users. Proceedings of the Fourteenth Conference on Uncertainty in Ar.ficial Intelligence, July 1998.
• Towards a Bayesian model for keyhole plan recogni>on in large domains Albrecht, Zukermann, Nicholson, Bud
Models rela>onship between user’s recent ac>ons and goals (help needs)
Probabilis>c goal persistence
Human-‐aware Robo.cs Cogni>ve mode { normal, error }
Dynamic Bayesian Nets
GPS reading
Edge, velocity, posi>on
Data (edge) associa>on
Transporta>on mode
Trip segment
Goal
zk-‐1 zk
xk-‐1 xk
θk-‐1 θk
Time k-1 Time k
mk-‐1 mk
tk-‐1 tk
gk-‐1 gk
ck-‐1 ck
Learning and Inferring Transportation Routines Lin Liao, Dieter Fox, and Henry Kautz, Nineteenth National Conference on Artificial Intelligence, San Jose, CA, 2004.
Human-‐aware Robo.cs Assisted cogni>on
l Understanding human behavior from low-‐level sensory data l Using commonsense knowledge l Learning individual user models
l Ac>vely offering prompts and other forms of help as needed
l Aler>ng human caregivers when necessary hpp://www.cs.washington.edu/assistcog/
Computer systems that improve the independence and safety of people suffering from cognitive limitations by…
Human-‐aware Robo.cs Ac>vity Compass • Zero-‐configura>on personal
guidance system – Learns model of user’s travel
on foot, by public transit, by bike, by car
– Predicts user’s next des>na>on, offers proac>ve help if lost or late
• Integrates user data with external constraints – Maps, bus schedules,
calendars, … – EM approach to clustering &
segmen>ng data
The Ac>vity Compass Don Paperson, Oren Etzioni, and Henry Kautz (2003)
Human-‐aware Robo.cs Ac>vity of daily living monitor & prompter
Founda>ons of Assisted Cogni>on Systems. Kautz, Etzioni, Fox, Weld, and Shastri, 2003
Human-‐aware Robo.cs Recognizing unexpected events using online model selec>on
• User errors, abnormal behavior
• Select model that maximizes likelihood of data: – Generic model – User-‐specific model – Corrupt (impaired) user
model • Neurologically-‐plausible
corrup>ons – Repe>>on – Subs>tu>on – Stalling
Fox, Kautz, & Shastri
fill keple put keple on stove
fill keple put keple on stove
put keple in closet