Enabling collaboration with intention inference using Partially Observable Markov Decision Processes

Abstract

Efficient collaboration between two humans jointly solving a task is often only possible if both humans know the other’s intention. They could communicate their intention verbally or simply observe one another until they have reasoned about the other’s intention before selecting their own approach to completing the task. This is relatively easy for humans to do due to their Theory of Mind - their ability to reason about other’s behaviour to develop a suitable response. Imbuing a robot with the same ability to enable collaboration with a human on those same tasks is complex as the robot lacks a Theory of Mind and so this system needs to be developed to allow the robot to infer a human’s intention. In this dissertation, a reasoning mechanism akin to Theory of Mind for a robot to reason about intentions is developed called an Action Trajectory Encapsulated Partially Observable Markov Decision Process (ATE-POMDP). This is modelled from the action trajectories of the human, whereby the agent aims to infer the current intention of the human based on observing their actions. Once their intention is found, the robot selects a task to complete itself, thereby collaborating with the human by completing some appropriate sub-components of the task. The ATE-POMDP is tested in simulation with simulated humans against other collaborative teams in gridworld environments to establish the effectiveness of this approach to intention inference. The results show that the ATE-POMDPs provide performance comparable to the other teams when there is a small collision time cost but are the best option for intention inference when collision time costs are large. This makes using ATE-POMDPs suitable to enable collaboration through intention inference