CSE 591: Human-aware Robotics

Transcription

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 Nov 8, 2016 Slides adapted from Subbarao Kambhampati This set of slides borrows from various online sources; it is used for educational purposes only.

2 Challenges in human-aware robotics Perception of humans Human recognition, human tracking, and activity recognition Human-robot interface Command recognition, gesture recognition Modeling of humans Goal and intent recognition, human decision and behavioral models, expectation, model learning Human-aware decision making Human-aware planning, reinforcement learning and inverse reinforcement learning.

3 Human-aware Decision Making Human-aware planner Human modeling Plan generation Robot models Human teammate Observations

4 Planning: The Canonical View A fully specified problem --Initial state --Goals (each non-negotiable) --Complete Action Model The Plan

5 Modeling of Humans When there are humans in the loop Capabilities Goal/plan preferences Goal/plan recognition Violated Assumption: Assumptions: Complete Models àcomplete àcomplete Action Action Descriptions Descriptions (fallible domain writers) àfully àfully Specified Specified Preferences Preferences (uncertain users) àpackaged àall objects planning problem the world (Plan known Recognition) up front àone-shot àone-shot planning planning (continual revision) Planning Allows is no planning longer to a pure be a pure inference inference problem problem! L The humans in the loop can ruin a really a perfect day L Traditional Planning Underlying System Dynamics

6 Planning: The Canonical View A fully specified problem --Initial state --Goals (each non-negotiable) --Complete Action Model The Plan

7 Human-aware Decision Making Human-aware planner Human modeling Plan generation Robot models Human teammate Observations?

8 Challenges in Human-Aware Planning & Decision Making Interpret what humans are doing Plan/goal/intent/preference/capability recognition Plan with incomplete domain models Robust planning with lite models (Learn to improve domain models) Continual planning/replanning Commitment sensitive to ensure coherent interaction Explanations/Excuses Excuse generation can be modeled as the (conjugate of) planning problem Asking for help/elaboration Reason about the information value

9

10 Mixed-initiative planner Decision support systems

11 Decision Theoretic Assistance Model world dynamics as a Markov decision process (MDP) Model user as a stochastic policy G U t P(G) P(U t G, W t ) A t Goal Distribution Awareness, interaction or teaming? Action distribution conditioned on goal and world state Transition Model W t W t+1 P(W t+1 W t, U t, A t ) U 1 W 1 W 2 W 3 W 4 A 1 U 2? Given: model, action sequence Output: assistant action

12 Mixed-initiative planner Ø Decision support systems HRT, virtual assistant Implicit communication HRT

13 Decision Theoretic Assistance Model world dynamics as a Markov decision process (MDP) Model user as a stochastic policy G U t P(G) P(U t G, W t ) A t Goal Distribution Direct plan structure, Custom interface, NL, pre-specified constraints, implicit? Action distribution conditioned on goal and world state Transition Model W t W t+1 P(W t+1 W t, U t, A t ) U 1 W 1 W 2 W 3 W 4 A 1 U 2? Given: model, action sequence Output: assistant action

14 Decision support systems: Critiques, subgoals &capabilities), state, belief, excuses, explanations... Implicit behavior HRT

15 Decision Theoretic Assistance Model world dynamics as a Markov decision process (MDP) Model user as a stochastic policy G U t P(G) P(U t G, W t ) A t Goal Distribution Goal, plan, model, Implicit behavior? Action distribution conditioned on goal and world state Transition Model W t W t+1 P(W t+1 W t, U t, A t ) U 1 W 1 W 2 W 3 W 4 A 1 U 2? Given: model, action sequence Output: assistant action

16

17 Decision Theoretic Assistance Model world dynamics as a Markov decision process (MDP) Model user as a stochastic policy G U t P(G) P(U t G, W t ) A t Goal Distribution Incomplete preference, dynamics? Action distribution conditioned on goal and world state Transition Model W t W t+1 P(W t+1 W t, U t, A t ) U 1 W 1 W 2 W 3 W 4 A 1 U 2? Given: model, action sequence Output: assistant action

18 Human aware robot planning implicit Interaction constraints Decision theoretic Assistance

19 Human aware robot planning implicit Interaction constraints A few more examples

20 Case studies: Ø Goal uncertainty Ø Plan uncertainty Ø Proactive help

21

22 room1 room5

23

24 Case studies: Ø Goal uncertainty Ø Plan uncertainty Avoid resource conflict Proactive help

25 Stigmergic Collaboration in human robot cohabitation The robot coordinates it s own behavior to suit the human s predicted plans to minimize conflicts e.g. planning with conflicts on shared resources [ICAPS PlanRob 2015] The robot directly interacts with the human s plans to assist/coordinate by making positive interventions e.g. planning for serendipity [IROS 2015] Much of the planning challenge is about defining the interac5on constraints under which the robot s planning process takes place. à The interac5on constraints themselves are informed by the plan(s) of the human (as recognized by the robot)

26 Case studies: Ø Goal uncertainty Ø Plan uncertainty Avoid resource conflict Proactive help

27 Planning with Resource Conflicts Overview & System Components Informa5on from the predicted plans concisely represented as resource profiles and fed to the planning stage.

28 Current Use Case Urban Search and Rescue (USAR) scenario Commander can perform triage (needs to get a medkit to do so) The Robot can also conduct triage or deliver medkits if requested The medkits are the shared resources here the robot must deconflict its plans to use the medkit with that of the human s.

29 29 Resource Profiles different levels of abstraction We can have profiles at different levels of abstraction to reason about different aspects of the plan Yes/no of resource usage Profiles over actual groundings of the resource variables

30 An Integer-Programming based Planner Modelling Constraints Objective function minimizes cost of plan, overlap of usage profiles and maximizes success rate Standard state equations, add and delete effects Planning with Stochastic Resource Profiles: An Application to Human-Robot Cohabitation. Tathagata Chakraborti, Yu Zhang, David Smith, Subbarao Kambhampati

31 31 The Planner Modelling Constraints Produce resource usage profiles for robot s plan Planning with Stochastic Resource Profiles: An Application to Human-Robot Cohabitation. Tathagata Chakraborti, Yu Zhang, David Smith, Subbarao Kambhampati

32 32 Adding Communication No good plans Having the ability to communicate changes the dynamics of the situation considerably the robot can now ask to use a resource during a specified period of time. Particularly useful if plans are too costly for the robot, or their success probabilities are too low, i.e. there exists no plan with zero conflicts In a non-teaming scenario communication can be unwanted overhead the profiles minimize this by telling the robot exactly what to communicate on which resources Planning with Stochastic Resource Profiles: An Application to Human-Robot Cohabitation. Tathagata Chakraborti, Yu Zhang, David Smith, Subbarao Kambhampati

33 Modeling Behavior 33 Compromise Robot settles for a suboptimal plan CommX has to do triage in room1, Robot is tasked to conduct triage in hall3 optimal plans require medkit1 from room2 for both agents.

34 Modeling Behavior 34 Opportunism Robot senses favourable turn of events CommX has to do triage in room1, Robot is tasked to conduct triage in hall3 optimal plans require medkit1 from room2 for both agents. When planning horizon is increased

35 Modeling Behavior 35 Negotiation Robot communicates to resolve conflict CommX has to do triage in room1, Robot is tasked to conduct triage in hall3 optimal plans require medkit1 from room2 for both agents.

36 Case studies: Ø Goal uncertainty Ø Plan uncertainty Avoid resource conflict Proactive help

37 Serendipitous Interactions 37 CommX has to conduct triage in room1. The robot fetches medkit2 from room3 and drops it off in hall3 before CommX passes by, thus saving him the effort to get a medkit himself.

38 Planning for Serendipity 38

39 39 Interaction Constraints Plan Interruptibility and Plan Preservation As we saw in the examples, the robot s planned intervention must adhere to a set of restrictions in order to be helpful to the human Constructing the composite plan from the individual plan Plan Interruptibility identify parts of the original individual plan that may be removed. Preservation Constraints given that the human is not expecting help, the rest of the human s plan should be executable as is. This means certain features of the original plan prefix and suffix needs to be preserved in the composite plan. Planning for Serendipity - Altruism in Human-Robot Cohabitation. Tathagata Chakraborti, Gordon Briggs, Kartik Talamadupula, Matthias Scheutz, David Smith, Subbarao Kambhampati

40 Human aware robot planning implicit Interaction constraints Summary of case studies

41 Challenges in Human-Aware Planning & Decision Making Interpret what humans are doing Plan/goal/intent/preference/capability recognition Plan with incomplete domain models Robust planning with lite models (Learn to improve domain models) Continual planning/replanning Commitment sensitive to ensure coherent interaction Explanations/Excuses Excuse generation can be modeled as the (conjugate of) planning problem Asking for help/elaboration Reason about the information value