Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Informative Communication of Robot Plans

Michele Persiani, Thomas Hellstrom

TL;DR

This paper addresses how robots should verbally communicate their plans to humans by accounting for the user's prior beliefs. It introduces an information-theoretic criterion, based on a second-order theory of mind, to select the most informative verbalizations that maximize Information Gain with respect to the robot's true model. The approach is implemented in PDDL as probabilistic Bayesian-network-based planning, and evaluated on the PUCRS dataset as well as a human-operator user study, demonstrating that informative verbalizations enable faster and more accurate goal understanding than incremental or purely order-based strategies. The findings suggest that communicating goal-related content early and choosing actions that discriminate the robot's internal state yields substantial efficiency gains in human-robot plan explanation and has practical implications for designing interpretable AI systems.

Abstract

When a robot is asked to verbalize its plan it can do it in many ways. For example, a seemingly natural strategy is incremental, where the robot verbalizes its planned actions in plan order. However, an important aspect of this type of strategy is that it misses considerations on what is effectively informative to communicate, because not considering what the user knows prior to explanations. In this paper we propose a verbalization strategy to communicate robot plans informatively, by measuring the information gain that verbalizations have against a second-order theory of mind of the user capturing his prior knowledge on the robot. As shown in our experiments, this strategy allows to understand the robot's goal much quicker than by using strategies such as increasing or decreasing plan order. In addition, following our formulation we hint to what is informative and why when a robot communicates its plan.

Informative Communication of Robot Plans

TL;DR

This paper addresses how robots should verbally communicate their plans to humans by accounting for the user's prior beliefs. It introduces an information-theoretic criterion, based on a second-order theory of mind, to select the most informative verbalizations that maximize Information Gain with respect to the robot's true model. The approach is implemented in PDDL as probabilistic Bayesian-network-based planning, and evaluated on the PUCRS dataset as well as a human-operator user study, demonstrating that informative verbalizations enable faster and more accurate goal understanding than incremental or purely order-based strategies. The findings suggest that communicating goal-related content early and choosing actions that discriminate the robot's internal state yields substantial efficiency gains in human-robot plan explanation and has practical implications for designing interpretable AI systems.

Abstract

When a robot is asked to verbalize its plan it can do it in many ways. For example, a seemingly natural strategy is incremental, where the robot verbalizes its planned actions in plan order. However, an important aspect of this type of strategy is that it misses considerations on what is effectively informative to communicate, because not considering what the user knows prior to explanations. In this paper we propose a verbalization strategy to communicate robot plans informatively, by measuring the information gain that verbalizations have against a second-order theory of mind of the user capturing his prior knowledge on the robot. As shown in our experiments, this strategy allows to understand the robot's goal much quicker than by using strategies such as increasing or decreasing plan order. In addition, following our formulation we hint to what is informative and why when a robot communicates its plan.

Paper Structure

This paper contains 10 sections, 3 equations, 3 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Background
  3. Method
  4. Informative Communication
  5. Implementation in PDDL
  6. Experiments
  7. Tests on the PUCRS dataset
  8. User Study
  9. Discussion
  10. Conclusion

Figures (3)

  • Figure 1: $P_R$: model of the robot working on task $T$. $P_R^H$: second-order theory of mind of the robot. The goal of interpretable behavior and verbalizations is to reduce the distance $|P_R - P_R^H|$. Here we consider the minimum amount of actions, and their order, that the robot should communicate such that $|P_R - P_R^H| < \theta$.
  • Figure 2: Robot model ($P_R$) and second-order theory of mind of the user ($P_R^H$) as equivalent Bayesian Networks. The cross-entropy $H$ measures the difference between robot's and user's estimated state of mind a posteriori of a verbalization $o$.
  • Figure 3: Entropy gains and distance-to-goal measures for the automated benchmarks.