Guiding users Clause Samples

Guiding users. Guiding a group is a complex problem, since the robot needs to adapt to its followers needs, which can be ambiguous. To solve this problem, we created a separate planning framework, that we called Collaboration Planners, based on hierarchical MOMDPs (Mixed Observability Markov Decision Pro- cess). A MOMDP models the decision process of an agent in situations where the result of an action is partly random, and can lead to several outcomes. In addition, in a MOMDP, the system state is split in an observable set and a hidden set, which cannot be fully observed and must be inferred from obser- vations received from the environment. MOMDPs are a variation of POMDPs (Partially Observable Markov Decision Process), where the system state is completely hidden. Partitioning the system state in a hidden and observable set simplifies the computation of a solution to the model, which is one of the main problems of POMDPs [3]. We use a hierarchical framework [5], where the system model is split into a main MOMDP module and several MOMDP sub-models, each one related to a different action. The models are solved separately, leading to the computation of different, simpler, policy functions. At run-time, the system interacts with the main module, providing values for the set of observations and for the observed variables, and receiving an action as result. Based on this action, the system will contact a different sub-model, receiving the final action to execute. Using hierarchical MOMDPs we can represent a set of models, with a greatly reduced complexity, and easily expand it if we want to implement new actions or to add more complex behaviors. The architecture of our system is shown in Figure 2 A).