Social Hierarchical Learning

Bradley HayesAAAI 2015 · DOI

Abstract

This paper presents a framework for "Social Hierarchical Learning" (SHL), a novel approach to human-robot collaboration (HRC) that leverages social signals to build hierarchical task models autonomously. By treating social cues as planning heuristics, the system enables robots to transition from passive execution to active supportive roles. The framework is validated through a furniture assembly task, demonstrating improved goal inference and the acquisition of supportive motor primitives that reduce human cognitive and physical load.

Introduction

As robots transition from industrial cages to shared workspaces, the challenge of human-robot collaboration shifts from simple execution to complex, real-time coordination. Effective HRC requires robots to perform state estimation, goal inference, and action planning simultaneously. However, most existing systems rely on "rich, hierarchical task models" provided upfront, which are difficult to extract from human demonstrations. Furthermore, robots often ignore the subtle non-verbal social cues humans use to communicate intent. This research addresses these bottlenecks by integrating social heuristics into task learning and planning, moving toward a "leader-assistant" partnership.

Methodology

The proposed framework bridges the gap between raw demonstration and fluid collaboration through a three-pronged approach:

  1. Learning Task Structure: Instead of passive observation, the robot employs active learning to guide instructors toward providing informative examples. These are used to construct flexible Hierarchical Task Networks (HTNs) derived from sub-goal ordering constraints within Semi-Markov Decision Processes (SMDPs).
  2. Collaborative Task-level Planning: The framework integrates human social signals into a multi-agent planner. It treats sequential motor planning as a coupled symbolic planning and geometric constraint satisfaction problem, allowing the robot to prioritize goals based on the human partner's perceived needs.
  3. Supportive Motor Primitives: The robot acquires actions specifically designed to assist humans, such as holding a piece of furniture steady. These are encoded as Partially Observable Markov Decision Processes (POMDPs) associated with specific states in a Hierarchical Hidden Markov Model (HHMM) used to track task progress.

Experimental Evaluation

The framework was evaluated using a furniture construction task, specifically chair assembly. This domain serves as a representative testbed for sequential motor tasks requiring high levels of coordination and manual dexterity.

Figure 1: Robot providing supportive stabilization (materials stabilization) during a furniture assembly task, allowing the human partner to focus on fastening components.

Key Results and Insights

  • Supportive Behavior Taxonomy: The research identifies five distinct categories of supportive robot behaviors: materials stabilization, materials retrieval, collaborative object manipulation, awareness improvement, and task progression guidance.
  • Multi-resolution Goal Inference: By utilizing HHMMs, the robot can infer human goals at varying levels of abstraction. This improves state estimation accuracy without the computational overhead associated with flat models.
  • Minimal Symbolic Requirements: The algorithm for building task hierarchies works with minimal prior symbolic knowledge, making it highly compatible with Learning from Demonstration (LfD) techniques.
  • Knowledge Transfer: The multi-level abstraction allows the robot to reuse high-level policies learned in one scenario for different future tasks, enhancing the system's versatility.

Conclusion and Future Work

The Social Hierarchical Learning framework provides a robust foundation for safer and more efficient HRC. While the current system focuses on non-verbal social signals, future work aims to integrate Natural Language Processing (NLP). By using autonomously constructed hierarchical task models as a grounding for language, robots could eventually utilize verbal communication as a reliable, real-time planning cue in dynamic and unpredictable environments.

Summary updated

BibTeX
@inproceedings{hayes2015social,
  author = {Bradley Hayes},
  title = {{Social Hierarchical Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9266},
}
Figure