No abstract

In multi-agent systems, the ability to infer intentions allows artificial agents to act proactively and with partial information. In this paper we propose an algorithm to infer a speakers intentions with natural language analysis combined with plan recognition. We define a Natural Language Understanding component to classify semantic roles from sentences into partially instantiated actions, that are interpreted as the intention of the speaker. These actions are grounded to arbitrary, hand-defined task domains. Intent recognition with partial actions is statistically evaluated with several planning domains. We then define a Human-Robot Interaction setting where both utterance classification and plan recognition are tested using a Pepper robot. We further address the issue of missing parameters in declared intentions and robot commands by leveraging the Principle of Rational Action, which is embedded in the plan recognition phase.

No abstract

This document contains a preprint version of the paper Policy Regularization for Legible Behavior that has been submitted to the Topical Collection on Human-aligned Reinforcement Learning for Autonomous Agents and Robots at the Springer journal on Neural Computing and Applications (NCAA) 2021. The contents of this document are subject to change, please refer to the NCAA version of the paper if available.

In this paper we propose a method to augment a Reinforcement Learning agent with legibility. This method is inspired by the literature in Explainable Planning and allows to regularize the agent’s policy after training, and without requiring to modify its learning algorithm. This is achieved by evaluating how the agent’s optimal policy may produce observations that would make an observer model to infer a wrong policy. In our formulation, the decision boundary introduced by legibility impacts the states in which the agent’s policy returns an action that is non-legible because having high likelihood also in other policies. In these cases, a trade-off between such action, and legible/sub-optimal action is made. We tested our method in a grid-world environment highlighting how legibility impacts the agent’s optimal policy, and gathered both quantitative and qualitative results. In addition, we discuss how the proposed regularization generalizes over methods functioning with goal-driven policies, because applicable to general policies of which goal-driven policies are a special case.