Towards Computational Modeling of Human Goal Recognition.

Area of Science:

• Human-Robot Interaction
• Cognitive Science
• Artificial Intelligence

Background:

• Emergence of rationality-based plan and goal-recognition algorithms in AI.
• These algorithms dynamically generate plans, avoiding large plan libraries.
• Prior research indicates human recognition of robot actions can be faster for non-optimal, less rational movements.

Purpose of the Study:

• To evaluate existing rationality-based plan-recognition algorithms against human decision-making data.
• To investigate the hypothesis that humans use plan recognition to infer goals.
• To develop a novel algorithm that better explains human observational choices in human-robot collaboration.

Main Methods:

• Experimentation with various rationality-based recognition algorithms on existing human-robot collaboration data.
• Development of a novel offline recognition algorithm integrating plan-library and rationality-based approaches.
• Comparison of the novel algorithm's performance against existing methods using the same dataset.

Main Results:

• Existing literature algorithms failed to account for human subject decisions in recognizing robot actions.
• The novel proposed algorithm demonstrated a significantly better fit to the experimental data compared to current methods.
• The new algorithm successfully combines elements of both rationality-based and plan-library based recognition strategies.

Conclusions:

• Current rationality-based plan-recognition algorithms do not accurately model human decision-making processes.
• Human plan recognition may serve as a precursor to goal recognition, integrating observations with known plans.
• The novel hybrid algorithm shows promise for more accurately simulating human-like recognition in human-robot interaction.