Two Heads Are Better Than One: A Dimension Space for Unifying Human and Artificial Intelligence in Shared Control

Document Title

Two Heads Are Better Than One: A Dimension Space for Unifying Human and Artificial Intelligence in Shared Control

Document Information

• Subject Area: Shared control in the fields of artificial intelligence and human-computer interaction
• Keywords: shared control, human-machine collaboration, design space analysis, artificial intelligence, collaborative interaction, supervisory control, assistive technology, AI roles

Research Background and Issues

• Problem or Challenge:

  • Shared control is a novel interaction paradigm that combines artificial intelligence (AI) with human intelligence and is widely applied across various domains. However, there is a lack of a unified language and design framework to describe and compare different systems across these domains.
  • Different fields (e.g., mobility assistance, driving, surgery, and gaming) employ diverse design concepts and terminologies for shared control, hindering cross-domain knowledge sharing and innovation.
  • Shared control requires a tool to help understand the multitude of possible design approaches, identify design gaps, and conceive new solutions.

• Significance and Importance:

  • The potential of shared control lies in enhancing system accessibility, safety, precision, and creativity for human users.
  • A unified approach can assist designers in better comparing designs, identifying design patterns, and fostering future innovations.

• Research Motivation and Related Work:

  • Building on the influential design space analysis methodology and foundational theories of shared control, the authors propose a multidimensional design space model to address the current gaps in shared control between humans and AI.
  • This study selected 55 shared control systems from various fields as a dataset and systematically summarized and analyzed their design characteristics.

Solution

• Proposed Solution:

  • The authors designed a dimension space model for shared control, comprising four axes: AI Role, Supervision, Influence, and Mediation.
  • The primary goal of the dimension space is to provide a layered analytical tool for shared control systems, enabling designers to describe existing systems, identify high-level design patterns, and develop new design concepts.

• Innovations:

  • Introduced the first classification method for shared control centered on "design space analysis."
  • Summarized high-level design patterns for shared control across domains, identifying and addressing design gaps between fields.
  • Utilized intuitive "Kiviat diagrams" and related tools to represent complex design variables as visually comparable patterns.

• Implementation Steps and Techniques:

  1. Systematic Induction and Classification: Classified 55 shared control systems and constructed the dimension space based on design characteristics.
  2. Definition of Four Core Dimensions:
     • AI Role: Distribution of AI roles in support, delegation, takeover, and complementarity.
     • Supervision: Describes the supervisory relationship between humans and AI (e.g., AI supervising humans, humans supervising AI, mutual supervision).
     • Influence: Explains the degree of mutual influence between humans and AI, such as independent, explanatory, guiding, and collaborative.
     • Mediation: Describes how human and AI control signals are integrated, such as "merging" or "selection."
  3. Design Pattern Summary: Identified common design patterns and described the interaction principles underlying these systems.

Research Outcomes

• Specific Outcomes:

  • Developed a comprehensive dimension space model for shared control, providing a classification and comparison tool for cross-domain designers.
  • Proposed and analyzed design patterns in six specific fields (e.g., mobility assistance, gaming, and surgical shared control patterns).
  • Identified common design patterns (e.g., "Alert Assistant," "Supportive Guardian," "Coordinated Negotiator," and "Equal Partner") and their applicability across different domains.

• Comparison with Existing Solutions:

  • Compared to existing models, this model focuses more on analyzing user experience and human-computer interaction methods in design, rather than solely on technology or control logic.
  • Provides a shared platform for interdisciplinary communication, enabling designers to apply experiences from one field to others.

• Experimental or Evaluation Results:

  • Analysis of 55 systems revealed many similarities among independently developed shared control systems across different domains, confirming the applicability and explanatory power of the dimension space.
  • Established design patterns common across multiple domains, laying the foundation for exploring new application possibilities.

• Limitations and Future Directions:

  • Limitations:
    • The current model is primarily based on an existing dataset and requires validation with systems from more fields.
    • Lacks real-world user experiments to evaluate the advantages and disadvantages of different design patterns in actual interactions.
  • Future Directions:
    • Expand the model's applicability while developing automated tools to help designers quickly classify and generate new designs.
    • Propose more refined design recommendations for specific fields, such as sustainable driving or smart city design.
    • Develop new interaction prototypes to validate user acceptance and effectiveness of normative patterns.