Probabilistic Human Intent Recognition for Shared Autonomy in Assistive Robotics.

Area of Science:

• Human-Robot Interaction
• Artificial Intelligence
• Robotics

Background:

• Effective human-robot collaboration in shared autonomy hinges on accurately predicting human intentions.
• Meaningful assistance requires the autonomy to infer the human collaborator's intended goal.
• Existing methods often struggle with diverse control interfaces and personalized user behavior.

Purpose of the Study:

• To develop a mathematical formulation for intent inference in assistive teleoperation under shared autonomy.
• To probabilistically infer user goals by fusing non-verbal observations and modeling human behavior.
• To personalize intent recognition through user-customized adjustable rationality.

Main Methods:

• A recursive Bayesian filtering approach is employed to probabilistically reason about user goals.
• Multiple non-verbal observations and contextual information are fused for intent recognition.
• Human agent behavior is modeled as goal-directed actions with adjustable rationality, optimized per user.

Main Results:

• The proposed approach demonstrates superior or comparable performance to existing methods in intent inference across various scenarios and tasks.
• Human subject studies validate the effectiveness of probabilistic modeling and user-customized adjustable rationality.
• Analysis reveals the impact of control interface limitations on intent inference accuracy.

Conclusions:

• Probabilistic modeling and incorporating goal-directed human behavior with user-customized rationality significantly benefit intent inference.
• The developed intent inference approach directly enhances shared autonomy performance.
• Control interface characteristics are critical factors influencing the success of intent inference in assistive teleoperation.