Human-Aware Planning Revisited : A Tale of Three Models

Human-aware planning requires an agent to be aware of the mental model of the humans, in addition to their physical or capability model. This not only allows an agent to envisage the desired roles of the human in a joint plan but also anticipate how its plan will be perceived by the latter. The human mental model becomes especially useful in the context of an explainable planning (XAIP) agent since an explanatory process cannot be a soliloquy, i.e. it must incorporate the human’s beliefs and expectations of the planner. In this paper, we survey our recent efforts in this direction. Cognitive AI teaming (Chakraborti et al. 2017a) requires a planner to perform argumentation over a set of models during the plan generation process. This is illustrated in Figure 1. Here, M is the model of the agent embodying the planner (e.g. a robot), and M is the model of the human in the loop. Further,Mh is the model the human thinks the robot has, andMr is the model that the robot thinks the human has. Finally, M̃h is the robot’s approximation ofMh ; for the rest of the paper we will be using Mh to refer to both since, for all intents and purposes, this is all the robot has access to. Note that the human mental modelMh is in addition to the (robot’s belief of the) human modelMr traditionally encountered in human-robot teaming (HRT) settings and is, in essence, the fundamental thesis of the recent works on plan explanations (Chakraborti et al. 2017b) and explicable planning (Zhang et al. 2017). The need for explicable planning or plan explanations occurs when the models – M and Mh – diverge so that the optimal plans in the respective models may not be the same and hence optimal behavior of the robot in its own model is inexplicable to the human. This is also true for discrepancies betweenM and Mr when the robot might reveal unrealistic expectations of the human in a joint plan. An explainable planning (XAIP) agent (Fox et al. 2017; Langley et al. 2017; Weld and Bansal 2018) should be able to able to deal with such model differences and participate in explanatory dialog with the human such that both of them can be on the same page during a collaborative activity. This is referred to as model reconciliation (Chakraborti et al. 2017b) and forms the core of the explanatory process of an XAIP agent. In this paper, we look at the scope of problems engendered by this multi-model setting and describe Figure 1: Argumentation over multiple models during the deliberative process of a human-aware planner (e.g. robot). the recent work in this direction. Specifically – We outline the scope of behaviors engendered by humanaware planning, including joint planning as studied in teaming using the human model, as well as explicable planning with the human mental model; We situate the plan explanation problem in the context of perceived inexplicability of the robot’s plans or behaviors due to differences in these models; We discuss how the plan explanation process can be seen as one of model reconciliation whereMh (and/or Mr ) is brought closer toM (M ); We discuss how explicability and explanation costs can be traded off during plan generation; We discuss how this process can be adapted to handle uncertainty or multiple humans in the loop; We discuss results of a user study that testify to the usefulness of the model reconciliation process; We point to ongoing work in the space of abstractions and deception using the human mental model.