Social planning for social HRI

Liz Sonenberg, Tim Miller, Adrian R. Pearce, Paolo Felli, Christian J. Muise, Frank DignumarXiv 2016 · arXiv

Abstract

For robots to be truly "social," they must move beyond simple rule-following and engage in social planning: the ability to reason about the beliefs, intentions, and potential misunderstandings of the humans they interact with. This paper proposes a framework where agents do not just plan for physical goals, but for social outcomes. By leveraging epistemic planning and "Theory of Mind" (ToM), the authors outline a layered deliberation architecture designed to mitigate "social brittleness" in human-robot interaction (HRI).

Introduction

Traditional AI planning is often technically "autistic," focusing on a single agent achieving a physical goal in a static or predictable environment. However, HRI is inherently multi-agent and social. A critical gap exists in current architectures: they struggle to account for the mental states of others. In a social setting, an action is more than a physical movement; it is a signal that changes what others believe.

Without the ability to reason about nested beliefs (e.g., "I believe that you believe that I know the way"), robots fail when human expectations diverge from the robot's internal model. The core challenge of social planning is managing the combinatorial explosion of possibilities when reasoning about these mental states while maintaining real-time responsiveness.

Proposed Framework

The authors advocate for a layered deliberation architecture rooted in the Belief-Desire-Intention (BDI) model, which balances reactive behaviors with high-level deliberation. The framework integrates several key concepts:

  • Epistemic Planning: Planning based on the knowledge and beliefs of all agents involved, allowing the robot to consider how its actions will change the mental states of others.
  • Theory of Mind (ToM): The computational ability to represent and reason about the perspectives and reasoning processes of other participants.
  • Grounding: A mechanism for establishing "common ground" to ensure all parties have a shared understanding of the current state and goals.
  • Stereotypes: To avoid the computational cost of deep ToM reasoning for every encounter, the authors suggest using stereotypical models of groups (e.g., "drivers," "pedestrians") as heuristic shortcuts.

Evaluation Scenarios

Rather than using standard training datasets, the paper proposes three "Challenge Scenarios" to evaluate the efficacy of social planning:

  1. The Intersection: A self-driving car and a pedestrian must negotiate right-of-way using social cues like eye contact or head nods.
  2. The Wumpus Hunt: A strategic scenario involving a Lord, a Peasant, and a Knight, where success depends on managing a misunderstanding about the Knight's identity.
  3. Hattari: A multi-player board game of deception and bluffing that requires players to detect lies and manipulate the information available to others.
The setup for the board game Hattari, used as a challenge scenario for social planning.

Key Insights and Results

The paper highlights that every physical action in a social context has a "social effect." For example, a robot stopping at a crosswalk is not just braking; it is signaling the intent to let a pedestrian pass. Managing these social motivations alongside practical goals requires finding a "sweet spot" between the expressivity of a social model and computational efficiency.

As a position paper, the results focus on conceptual progress and preliminary implementations. The authors report:

  • Advancements in efficient mechanisms for run-time plan generation in multi-actor contexts.
  • The development of novel mechanisms for stereotypical and empathetic reasoning.
  • A limited implementation of an artificial player for Hattari, demonstrating the use of epistemic planning to handle incomplete and manipulated information.

Limitations and Future Work

Achieving human-level social AI remains a significant challenge. Current hurdles include the "perceptual gap," where technology struggles to reliably interpret subtle human social cues like body language in real-time. Furthermore, there is a lack of formal metrics for what makes a social interaction "plausibly effective." Future research must focus on "social repair"—how agents can detect when a social interaction has failed and work to restore common ground.

Summary updated

Abstract

Making a computational agent 'social' has implications for how it perceives itself and the environment in which it is situated, including the ability to recognise the behaviours of others. We point to recent work on social planning, i.e. planning in settings where the social context is relevant in the assessment of the beliefs and capabilities of others, and in making appropriate choices of what to do next.
BibTeX
@misc{sonenberg2016social,
  author = {Liz Sonenberg and Tim Miller and Adrian R. Pearce and Paolo Felli and Christian J. Muise and Frank Dignum},
  title = {{Social planning for social HRI}},
  year = {2016},
  eprint = {1602.06483},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1602.06483},
}
Figure